Enhanced detonators detection in X

Scientific Reports volume 13, Article number: 14262 (2023) Cite this article

Metrics details

Detecting detonators is a challenging task because they can be easily mis-classified as being a harmless organic mass, especially in high baggage throughput scenarios. Of particular interest is the focus on automated security X-ray analysis for detonators detection. The complex security scenarios require increasingly advanced combinations of computer-assisted vision. We propose an extensive set of experiments to evaluate the ability of Convolutional Neural Network (CNN) models to detect detonators, when the quality of the input images has been altered through manipulation. We leverage recent advances in the field of wavelet transforms and established CNN architectures—as both of these can be used for object detection. Various methods of image manipulation are used and further, the performance of detection is evaluated. Both raw X-ray images and manipulated images with the Contrast Limited Adaptive Histogram Equalization (CLAHE), wavelet transform-based methods and the mixed CLAHE RGB-wavelet method were analyzed. The results showed that a significant number of operations, such as: edges enhancements, altered color information or different frequency components provided by wavelet transforms, can be used to differentiate between almost similar features. It was found that the wavelet-based CNN achieved the higher detection performance. Overall, this performance illustrates the potential for a combined use of the manipulation methods and deep CNNs for airport security applications.

The detection of dangerous objects in X-ray images of baggage has become important, particularly due to rising crime rates1. The performance of screening devices is strongly influenced by the target visibility, image display technology, and the security officers’ knowledge. However, visual inspection of these images is highly challenging due to the low prevalence of targets, variability in target visibility (resulting in lack of precision in object shape), overlapping objects, poor contrast that obscures image details, and the potential for causing false alarms2,3. Furthermore, the constant and repetitive nature of the task i.e., the security officers constantly looking at screens and frequently encountering the same types of detected objects, can lead to attention fatigue and impaired judgment4.

The most dangerous prohibited articles in passenger baggage are so called improvised explosive devices. Detecting the detonator of a bomb can be a challenge even for well-trained security officers. To address these issues, numerous algorithms and techniques have been developed to improve the quality of 2D radiographic images5,6,7,8,9,10,11,12,13. The Bag-of-Visual-Words (BoVW) detection technique, which is based on natural language processing and information retrieval, employs a statistical process for object detection and classification6. This technique has been successfully applied for the detection of explosives. It was used together with various other methods, including supervised feature learning by autoencoders approach7, K-Nearest Neighbors, Logistic Regression8, and Decision Trees9. BoVW was also employed for the detection of guns, shuriken or razor blades. These techniques are based on dictionaries formed for each class and the detection consists of Scale Invariant Feature Transform (SIFT) feature descriptors of randomly cropped image patches10. The BoVW model correlated to the Speeded up Robust Features (SURF) descriptor and Support Vector Machine (SVM) classifier was used for firearm detection, achieving an optimal true positive rate of 99.07% at a false positive rate of 20%11. Both random forest and SVM algorithms were used for firearm detection and a statistical accuracy of 94% was reported12. Single, two and multiple X-Ray views and four classifiers (i.e., Scale-Invariant Feature Transform, Oriented FAST and Rotated BRIEF, Binary Robust Invariant Scalable Keypoints and SURF) were considered to assess the performance of classification. A better performance of classification has been highlighted when a combination of two and multiple X-Ray views was considered13.

In recent years, CNNs have gained significant popularity in the field of X-ray image analysis for baggage screening14,15,16,17,18,19,20. The augmentation technique, a feature enhancement module and a multi-scale fused region of interest method enabled the development of new CNNs with more accurate and robust detection capabilities. Those CNNs have a significantly improved performance when dealing with densely cluttered backgrounds during the X-ray baggage inspection14. Various techniques were employed to overcome different shortcomings of deep CNNs which were caused by a shortage of training images. Thus, the transfer learning paradigm15,16, region-based CNN (R-CNN), mask-based CNN (Mask R-CNN) and detection architectures such as RetinaNet were used to provide object localization variants17,18 or to detect various items in the X-ray image of the baggage. In a similar fashion, the You Only Look Once (YOLO) architecture was used for X-ray images of baggage classification and for hazardous materials identification19. Moreover, an anchor-free CNN-based object detection method was proposed to address the problem of dangerous objects detection20.

Wavelet transforms are a popular tool in image denoising. They are mostly used in denoising operations without any prior knowledge of the noise model. Also, they are a useful tool for image enhancement. Apart from image denoising, which is a subjective process, image enhancement alters the image features to make it more appealing to the human eye21. Wavelet edges effects are noticeable in the processed images but wavelets are not currently used in X-ray security inspection. Implementation of the wavelet transforms for X-ray security inspection is still lacking even if they are extensively used in several contexts of chest X-ray images, due to their strong predictive capabilities. Various studies reported only the implementation of machine learning or deep learning radiomics models for predicting COVID-19 prognosis, based on edge detection or radiomic features extraction22,23,24. While the utilization of wavelet transforms with CNNs in medical image processing has been extensively studied, their application in other domains, such as aviation and transport security, has been relatively limited.

In the case of X-ray airport security data, large amounts of training data are not always available and collecting X-ray data with the special request of data annotation is very expensive. A study devoted to the performance of screeners’ assessment in detecting bomb detonators with 2D and 3D imaging was conducted in Ref.25. Despite the lower image quality in 3D imaging, this study found that the performance was almost similar to that of 2D imaging. In another approach, an Unsharp masking USM + CLAHE algorithm to process radiographic images for airport security was developed to effectively overcome the color distortion induced by the CLAHE image enhancement26. Generally, the majority of the previous studies have primarily focused on the detection of secondary explosive devices (such as C4, TNT, etc.) in X-ray images during baggage inspection. Nowadays, there are limited studies that specifically address the detection of detonators in the X-ray baggage inspection process. For instance, a Dual-Energy method was used to detect dangerous objects, including detonators, by differentiating between organic and inorganic materials27. The majority of existing scientific publications dealing with this topic focus on various algorithms devoted to enhancing the image quality and to improve the detection performance of detonators28, but few concentrate on wavelet-based detonators detection in X-ray images.

The aim of this paper is to introduce a new and efficient scheme for detecting detonators in X-ray baggage images by comparing different image manipulation methods and by evaluating their impact on the predictive capabilities of the classification models. To overcome the underutilization of wavelets in X-ray security inspections, we introduce wavelets as a manipulation method, which can be used to obtain images with higher resolution and more defined details. This allowed us to gain insights into the validity of the manipulation processes and how they relate to the performance of detonators detection. The experiments are conducted using the High Tech Detection Systems (HTDS) database. In the proposed approach we have chosen and built a well-established CNN architecture which had achieved excellent performance in object classification and detection29. We have conducted an extensive ablation study to establish an optimal configuration model with good performance across the dataset30,31. Thus, we have experimented with different images corrupted by Gaussian and salt-and-pepper noise, various altered hyper-parameters and different layer structures. The proposed CNN architectures perform two stages of analysis: (i) detonators detection within the raw X-ray image using the deep CNN-based classifier, the TensorFlow and Keras libraries and, (ii) the same CNN classifier framework is used when the input image set is pre-processed using the following methods: the CLAHE algorithm, which operates independently on the RGB images and furthermore, also on the individual color channels, the wavelet transforms with the HH and HL sub-bands, and, a combination of CLAHE and RGB-wavelet transform techniques. The outputs are analyzed in terms of accuracy, precision, recall, F1-score, and classification.

Our novel contributions compared to other state-of-the-art approaches can be summarized as follows:

We proposed a multiscale approach by combining CLAHE, wavelet transforms, and RGB-wavelet transforms with CNNs to address the issue that different X-ray image quality factors can make the detection of the detonators a difficult task.

We conduct experiments on manipulated images to find the proper technique able to achieve the highest detection performance.

The custom CNN architecture proved compatible with various image manipulation techniques, being able to exploit the distinguishability between classes of baggage, with and without detonators inside.

The proposed methods of image manipulation emulate various technical specifications and assess the detection performance. Besides its practical relevance, a comparison of these manipulation methods is also of theoretical interest. The proposed study validates wavelets as a new framework for further studies in multivariate multiresolution analysis of X-ray screening of passenger baggage.

In our opinion, these experiments are at the proof-of concept level. We have tried to demonstrate that our idea could be turned into a reality. However, at this stage, only relatively limited datasets are available for sound training and testing operation.

In this paper, we have investigated the ways in which the manipulation of the quality of X-ray images can create a feature map which leads to an improved discrimination of dangerous materials, like detonators. The deep CNN architecture works with the TensorFlow and Keras libraries, based on the High Tech Detection Systems (HTDS) database. The deep CNN only has two classes, namely detonators and non-detonators. We used the augmentation techniques to increase the number of samples and to avoid overfitting. The final dataset consists of 15,115 samples of which 4535 samples were used for our tests. A random 70/30 training/testing splits of the dataset has been performed. Due to sample scarcity in the original dataset, we used fivefold cross-validation during the training of raw X-ray images. The same CNN architecture was set up for all experiments.

Table 1 presents the binary classification performance for the raw image dataset. A classification accuracy of 0.9808 is reported. All detonator spot predictions are correct.

To analyze the effects of the contrast enhancement and wavelet transformation on the detection efficiency, the performance of the classification for each proposed approach was examined. A visual representation of how the radiographic image has been improved by using the CLAHE method is shown in Fig. 1. The results of the CLAHE image enhancement algorithm are presented in Table 2.

CLAHE enhancement of the original radiographic image and its results for each color channel for an RGB image (a) original; (b) Red channel; (c) Green channel; (d) Blue channel; (e) RGB image.

The CLAHE method is working very well on the green and blue channels, as well as on RGB images. In both cases, it has achieved an accuracy of 1.00. These values are higher than the accuracy of the raw radiographic images (Table 1), with a difference of 1.92%. All detonator spot predictions are correct for the G and B channels and for the RGB images. While it may be unrealistic to expect 100% accuracy values, we note that the CLAHE can add the discrimination potential by improving the contrast and image quality and by reducing the loss of details in the image. However, the number of TP samples (images with detonator) is lower.

To examine the effects of decomposing an image into a set of wavelets and to analyze the improvements of the local spectral and temporal information extraction, the X-ray image decomposition using the Haar, Db2, Coif2, and Sym2 wavelet functions has been performed. The HL and HH details coefficients or sub-bands are of interest. The results are shown in Fig. 2 and details regarding the performance of classification are presented in Table 3.

Wavelet transformation of the original radiographic image and the obtained results for two sub-bands, HH and HL. Daubechies2 (Db2), Haar, Coiflet2 (Coif2), and Symlet2 (Sym2) wavelet functions are used for the transformation.

As it can be seen, the Coif2 function at the level HH achieved an accuracy of 0.995, which is slightly higher than Sym2 (0.994) and higher than Db2 and Haar (0.990 and 0.986). At the HL level, the Haar function resulted in an accuracy of 1.00, surpassing all the other functions: Db2 (0.999), Coif2 (0.998), and Sym2 (0.995).

Additionally, the data in Table 3 shows that the accuracy results provided by wavelet transforms are superior to the accuracy obtained when raw radiographic images are directly fed to the CNN (Table 1). It also shows that the Haar transform has the same performance as the color channel decomposition in terms of accuracy (100%). However, the number of true positive TP samples (images with detonators) is smaller in the case of the RGB channel decomposition, while the number of true negative TN sample is similar.

When both the CLAHE RGB and wavelet transformation methods are combined, the result of pre-processing is shown in Fig. 3. A summary of the performance of classification is given in Table 4.

Image preprocessing using the CLAHE RGB and wavelet transforms. The X-ray image in the RGB space is contrast enhanced using CLAHE. The result is decomposed through the instrumentality of the wavelet transform applied in the HH and HL sub-bands.

It can be observed that the Haar wavelet at the HH level achieved an accuracy of 0.999. This is the highest accuracy among the other mother wavelet functions, namely Db2 (0.999), Coif2 (0.997), and Sym2 (0.992). For the HL level, the Haar and Coif2 wavelets obtained an accuracy of 0.999, which is still the highest when compared to the Db2 (0.996) and Sym2 (0.995).

Generally, the accuracy of classification obtained using the CLAHE RGB-wavelet method is significantly higher that the accuracy obtained for raw X-ray images, but almost similar to the results provided by the wavelet decomposition. We emphasize once again, that although the performance metrics have very good values, they alone are not enough to indicate any enhancement in the detection of detonators. In the wavelet decomposition case, the number of TP samples (images with detonators) is higher than in the case of RGB channel decomposition or wavelet transform while the number of true negative (TN samples) is almost identical. There is an insignificant number of FN samples (i.e., type II error) in the classification. To summarize the findings in our study, a descriptive statistic of the accuracy is provided in in Fig. 4.

CNN average accuracy of classification for raw X-ray images, and as well as the accuracies of the proposed preprocessing models. The central lines indicate the median accuracy values and the boxes indicate the interquartile range. The whiskers indicate the smallest and largest values. Both the raw X-ray images and HL wavelet processed images have zero skew for the accuracy distribution. The rest of the methods show accuracies with a slightly negatively skewed distribution. Overall, we obtained the best performance with the Haar wavelet transform at HL sub-band along with the CLAHE & RGB image method.

Although many researchers have used classical methods for contrast enhancement, we introduce wavelets as a manipulation method to obtain images with higher resolution and more defined details. With this approach, we investigated the accuracy of detection when there are subtle differences in the image details, i.e., edges enhancements, altered color information or different frequency components provided by wavelet transforms. All of these could be used to differentiate between almost similar features in images. We have achieved improved classification results by exploiting high-level features provided by wavelet transforms. The first key result presented in this paper is that the presence of detonators can be effectively discriminated by traditional image processing methods such as wavelet decomposition or by combining contrast enhancement techniques and Haar wavelet transforms at the HL sub-band combined with deep CNN model. The use of the deep CNNs for detonators detection resulted in an improved recall (the ability to identify the positive class aka images with detonators) and precision (how accurate the positive predictions are).

Although deep CNNs have been a booster for images classification, we are aware that both the quality of an image, together with the max-pooling down-sampling method may dilute or remove some features that are important for classification. In order to determine the optimal architecture and configuration of our CNN model, an ablation study has been performed for a clear understanding of the model’s performance. The consequence of altering some components or hyper-parameters was a decrease in the performance of the model. The current network architecture provides an optimal performance with low computational complexity (see Tables 6 and 7). The proposed CNN uses a built-in mechanism to extract different high-level features and introduce spatial invariance. Also, the networks learning abilities are improved by its structure. The convolution layers result in different image features being extracted to a different extent and the pooling operation builds the ability to detect the same feature in different images. By feeding the same CNN architecture with different variants of the original image (i.e., various features are highlighted by manipulation) the net can be more optimally trained and its pattern recognition ability and image interpretation performance are thus improved.

The wavelet transforms ensure an effective Gaussian noise removal and they can extract the detail of images pretty well. The HH and HL components contain textural representations of the original image. It is worth noting that, when we used different frequency components and different features of the colour channels, more informative features were extracted and a better performance of detection was achieved. We were able to capture fine and coarse details of the image and improve image quality. The CLAHE method performs local histogram equalization pixel by pixel, thus improving both the contrast and image quality. On the other hand, the wavelet transform method generally recovers features from regions of interest and processes the image in pairs of pixels.

For dual X-ray images, the color information could be relevant to the detection. We have noticed that RGB images shows better results. Certain colors are indicative of the presence or absence of prohibited items and the model can learn these relationships. It's important to note that in some cases, converting RGB images to grayscale or to color channels can introduce biases and distortions. In this case, the enhancement method could fail to balance the areas with uneven illumination and some details became invisible as subtle color deviations appear.

It can be seen that, unlike the original images, images processed with the CLAHE RGB-Wavelet method (HH and HL sub-bands) and wavelet transforms at HH and HL sub-bands suffer from false negatives and from false positives. Images processed with the CLAHE RGB and Haar wavelet (HL sub-band) suffer from neither false positives nor false negatives. Our preliminary results show that the CLAHE & RGB image methods (Table 2), along with the Haar wavelet transform applied for the HL sub-band (Table 3) achieve a 100% true positive rate. This means that these methods are well-suited for detonators identification through the CNN approach. The differences (albeit very small) between the image manipulation results in detonators detection are explained as follows: the contrast enhancement works using the intensity map of pixels, reduces the loss of details and doesn’t modify the pixel position in the image. CLAHE applies color balancing through histogram adjustment. The contribution of the local spectral and temporal information to the edge extraction provided by the wavelet transforms is considered. Usually, the noise information exists in the high-frequency component and the low-frequency component contains the relevant information about the image. HL images balance both aspects and CLAHE better reduces the noise in the HL sub-band.

To summarize the findings in our study, we have observed that increasing the number of preprocessing tasks (we refer here to the CLAHE—RGB—Haar wavelet and the HH and HL sub-bands) does not necessarily lead to an increase in accuracy. These high-frequency HH and HL sub-bands saved most of the information so that the CNN could preserve more features and provide an accurate classification.

Table 5 reports the comparison of the accuracy values reported in our work and other classification approaches and the detection strategies based on deep CNN architectures, CLAHE enhancement and wavelet transforms. We compare our experimental results with some previous studies that experimented with the same methods but with different images type and databases (only accuracies). For fair comparisons, we conducted the experiments on detonator detection using two pre-trained models, EfficientNetV2B0 and AlexNet. Both are pre-trained using the ImageNet database, which contains millions of labeled images.

As shown in Table 5, although the AlexNet has a high accuracy (0.9945), it generates a number of false positives (16 on average). The EfficientNetV2B0 has the worst performance, while our network outperforms it regardless of how the images are manipulated. These results indicate the robustness of our approach.

Our experiments are at the proof-of-concept level, but they demonstrate that our idea could be turned into reality. However, at this stage, there are some limitations. The dataset was relatively small to train our CNN. Data augmentation was our method of choice to deal with overfitting, but we are also devoted to the entropic capacity of our CNN model, that is, the amount of information that the CNN model can store. The architecture of our CNN model allows storing a large amount of information, so its potential to be more accurate by leveraging more features was increased. Another limitation could rise from the cluttered nature of the X-ray image input dataset. This could cause the CNN to fail to detect the detonators. The proposed solution demonstrates that the CNN is an accurate tool as the FN samples occurrence is insignificant (Tables 3, 4). A more complex and complete study (more images, various sizes for baggage, features extraction specific to the organic sample, another neural network type etc.) will be carried out in the future. Also, it would be interesting to extend the research to multiclass classification, including other dangerous objects in X-ray images such as, TNT, C4, and PBX (plastic-bonded explosives).

In order to detect suspicious objects, such as detonators, in the 2D radiographic images of baggage from the HTDS database, the proposed block diagram in Fig. 5 outlines the experiments conducted.

Block diagram of the proposed method after the ablation study. HH (High-High) and HL (High-Low) are two sub-bands of decomposition in wavelet transform.

The database used in this study is sourced from High Tech Detection Systems (HTDS), a French company specialized in the sales and maintenance of high-tech security equipment used for passenger screening, baggage and vehicle security, and freight36. The initial dataset consists of 6,500 images divided into two types: class 0, which consists of 5,500 images with detonators and class 1, which consists of 1,000 images without detonators, as shown in Fig. 6.

Examples of X-ray images of baggage with multiple objects (top row) and without/with detonator inside (bottom row). The detail of a detonator is illustrated. Different materials shown in different colors on a pseudo-color image from dual-energy X-ray scans illustrate the 2D image low quality and resolution.

The dataset has been split into training (70%) and test sets (30%) such that each split has a similar class distribution. The test set contains samples never used for training. To balance the number of samples in both classes we have performed random rotations to each sample to augment the dataset. The final dataset consists of 15,115 samples. The performance of classification is assessed on the test dataset.

The CLAHE enhancement method is used to improve the image quality and contrast in the dual-energy radiographic images23. The contrast can be manipulated by modifying the intensity map of pixels. It optimizes the object identification, preserves details through morphological processing, and performs RGB (red, green, blue) conversion.

Discrete wavelet transforms can extract spectral features. A 2D wavelet transform passes the image through a low-pass and high-pass filter and approximate and detailed parts are provided. This multi-resolution framework generates four sub-bands, namely one approximation coefficient (LL) and three detail coefficients, horizontal (HL), vertical (LH), and diagonal (HH), respectively. The LL sub-band contains no edge and is not useful for our goal. Both the HL and LH sub-bands contain almost similar information and for the sake of efficiency, only the HL is investigated. To this end, we used the following mother functions: Haar, Daubechies2 (Db2), Symlet2 (Sym2), and Coiflet2 (Coif2)21,22. During the wavelet transforms, the input image is convolved with low-pass and high-pass filters and down sampled to obtain the wavelet sub-bands. Only the first level wavelet decomposition is used. Both the HL and HH images are passed through hidden convolution layers, then batch normalization allows faster execution and solves the issue of poor convergence. Finally, a softmax layer is used to categorize the dual-energy radiographic images. To assess the capability of the pre-processed images to be meaningful for the deep CNN classification task, we also combine the CLAHE—RGB conversion with HH and HL sub-band wavelet transformations, using the same mother functions.

The ablation study allows a clear understanding of the proposed model by analyzing the consequences of altering images and components of the CNN architecture. For fair comparisons, the raw and wavelet transformed images from the HTDS database were used. The results of the entire ablation study are recorded in the Tables 6 and 7. To evaluate the performance of the ablation process, the same metrics (accuracy, precision, recall and F1-score) were used.

It can be seen that the CNN + wavelet transform models, for all wavelet families, achieves better accuracy values and outperforms the CNN model which used raw images as input.

We performed extensive experiments to examine the strength of the proposed CNN model for the binary classification of dual-energy radiographic images into two classes: detonator and non-detonator. The CNN architecture is implemented using TensorFlow and Keras. TensorFlow is a free and open-source software library for data flow and differentiable programming through a variety of functions. It is very convenient and flexible for building the current deep learning models37. Keras is an open-source neural network library written in Python and runs on top of TensorFlow38. The deep CNN is entirely based on five deeply separable layers (i.e., three convolutional layers and two dense layers). Table 8 shows the proposed hyperparameters for the deep CNN architecture model, after the ablation study. The first convolutional layer has 16 units, the second layer has 32 units, and the third layer has 64 units. The Rectified Linear Unit (ReLU) activation function is used, as it is a non-linear function and has the advantage of avoiding backpropagation errors. The Adam (Adaptive Moment Estimation) optimizer is used with the default learning rate of 0.001 and 1e − 6 decay. The batch size of 32 is the optimal solution. The right number of epochs depends on the inherent complexity of the dataset. We had employed six values for the epoch (60, 80, 100, 120, 140, 160) in an attempt to improve our model. The number of epochs without improvement in loss function plateau was between 55 and 65, for all experiments. The time per epoch is between 72 and 76 s, with an average loss of 0.0677. We have determined that a good time to stop training is around epoch 60. The accuracy of the proposed CNN with 60 epochs has been chosen as the optimal option with less time and reduced losses. Dense layers are fully connected layers that transform one 1D feature vector into a classification vector. They are followed by the ReLU activation functions. The last dense layer does not have a specified activation function because the logit values are used and the classification is performed using a softmax activation function.

For baselines comparisons, we quantify our proposed CNNs in terms of effectiveness and efficiency with two commonly used models, EfficientNetV2B0 and AlexNet. All the networks are trained using the same datasets for fair comparisons.

The performance of the classification is evaluated comparing the average accuracy, precision, recall and F1-score, using same test dataset and proposed manipulation methods. These measures were calculated using the confusion matrix binary classification problems (i.e., a 2 × 2 matrix). The binary confusion matrix provides accurate insights into the model's performance and determines if the model can effectively discriminate between classes and identify specific prediction errors, which can lead to model improvements or adjustments. The confusion matrix includes the following elements: TP = true positives, TN = true negatives, FP = false positives, FN = false negatives. The performance metrics derived from this matric are: recall = [TP/(TP + FN)], precision = [TP/(TP + FP)], accuracy = [(TP + TN)/(TP + TN + FP + FN)]), F1-score = 2[(precision × recall)/(precision + recall)]. All results are averaged over the test dataset. The average accuracy of the detonator detection over all samples indicates how many of our detonator and non-detonator predictions are correct. The average recall indicates how many of the detonator samples are predicted as such and the average precision shows how many of our detonator predictions are correct.

The dataset is available upon request from High Tech Detection Systems (HTDS, https://www.htds.fr/). The data are not publicly available. The data that support the findings of this study are available from the first author, [LO], upon reasonable request.

Source code is made available at https://gitfront.io/r/user-3214156/oGn8DHKwkwzB/inspection-by-image/.

Pozzo, F. R. D. International and EU legal frameworks of aviation security. In Air Transport Security (eds. Szyliowicz, J. S., Zamparini, L.) 43–62 (2018). https://doi.org/10.4337/9781786435200.00011.

Partridge, T. et al. Enhanced detection of threat materials by dark-field X-ray imaging combined with deep neural networks. Nat. Commun. 13, 4651. https://doi.org/10.1038/s41467-022-32402-0 (2022).

Article ADS CAS PubMed PubMed Central Google Scholar

Vukadinovic, D., Anderson, D. X-ray baggage screening and AI, EUR 31123 EN, Publications Office of the European Union, Luxembourg, JRC129088 (2022). https://doi.org/10.2760/46363.

Bolfing A., Halbherr T., Schwaninger A. How image based factors and human factors contribute to threat detection performance in X-ray aviation security screening. In Lecture Notes in Computer Science (ed. HCI and Usability for Education and Work, Holzinger, A.) 419–438 (Springer, 2008). https://doi.org/10.1007/978-3-540-89350-9_30.

Mery, D., Pieringer, C. Computer Vision for X-Ray Testing: Imaging, Systems, Image Databases, and Algorithms (Springer International Publishing, 2021). https://doi.org/10.1007/978-3-030-56769-9.

Zhang, Y., Jin, R. & Zhou, Z.-H. Understanding bag-of-words model: A statistical framework. Int. J. Mach. Learn. Cybern. 1, 43–52. https://doi.org/10.1007/s13042-010-0001-0 (2010).

Article Google Scholar

Chouai, M., Merah, M., Sancho-Gómez, J.-L. & Mimi, M. Supervised feature learning by adversarial autoencoder approach for object classification in dual X-ray image of luggage. J. Intell. Manuf. 31, 1101–1112. https://doi.org/10.1007/s10845-019-01498-5 (2020).

Article Google Scholar

Larsen, K., Petersen, J. H., Budtz-Jørgensen, E. & Endahl, L. Interpreting parameters in the logistic regression model with random effects. Biometrics 56, 909–914. https://doi.org/10.1111/j.0006-341X.2000.00909.x (2000).

Article CAS PubMed MATH Google Scholar

Andrews, J. T. A., Jaccard, N., Rogers, T. W., Griffin, L. D. Representation-learning for anomaly detection in complex X-ray cargo imagery. In Anomaly Detection and Imaging with X-Rays (ADIX) II (SPIE, 2017). https://doi.org/10.1117/12.2261101.

Mery, D., Svec, E. & Arias, M. Object recognition in X-ray testing using adaptive sparse representations. J. Nondestr. Eval. 35, 45. https://doi.org/10.1007/s10921-016-0362-8 (2016).

Article Google Scholar

Turcsany D., Mouton A., Breckon T. P. Improving feature-based object recognition for X-ray baggage security screening using primed visual words. In Proceedings IEEE International Conference on Industrial Technology (ICIT), Cape Town, South Africa, 1140–1145 (2013). https://doi.org/10.1109/ICIT.2013.6505833.

Kundegorski, M. E., Akçay, S., Devereux, M., Mouton, A., Breckon, T. P. On using feature descriptors as visual words for object detection within X-ray baggage security screening. In Proceedings of the 7th International Conference on Imaging for Crime Detection and Prevention (ICDP 2016), vol. 12(6) (2016). https://doi.org/10.1049/ic.2016.0080.

Tuli, A., Bohra, R., Moghe, T., Chaturvedi, N., Mery, D. Automatic threat detection in single, Stereo (Two) and multi view X-ray images. In Proceedings of IEEE 17th India Council International Conference (INDICON), New Delhi, India 1–7 (2020). https://doi.org/10.1109/INDICON49873.2020.9342253.

Gu, B., Ge, R., Chen, Y., Luo, L. & Coatrieux, G. Automatic and robust object detection in X-ray baggage inspection using deep convolutional neural networks. IEEE Trans. Ind. Electron. 68, 10248–10257. https://doi.org/10.1109/TIE.2020.3026285 (2021).

Article Google Scholar

Akçay S., Kundegorski M. E., Devereux M., Breckon T. P. Transfer learning using convolutional neural networks for object classification within X-ray baggage security imagery. In Proceedings of IEEE International Conference on Image Processing (ICIP), 2016, Phoenix, AZ, USA 1057–1061 (2016). https://doi.org/10.1109/ICIP.2016.7532519.

Griffin, L. D., Caldwell, M., Andrews, J. T. A. & Bohler, H. Unexpected item in the bagging area: Anomaly detection in X-ray security images. IEEE Trans. Inf. Forensics Secur. 14, 1539–1553. https://doi.org/10.1109/TIFS.2018.2881700 (2019).

Article Google Scholar

Gaus Y. F. A., Bhowmik N., Akçay S., Guillén-Garcia P. M., Barker J. W. et al. Evaluation of a dual convolutional neural network architecture for object-wise anomaly detection in cluttered X-ray security imagery. In Proceedings of International Joint Conference on Neural Networks (IJCNN), Budapest, Hungary 1–8 (2019). https://doi.org/10.1109/IJCNN.2019.8851829.

Subramani, M., Rajaduari, K., Choudhury, S. D., Topkar, A. & Ponnusamy, V. Evaluating one stage detector architecture of convolutional neural network for threat object detection using X-ray baggage security imaging. Revue d’Intelligence Artificielle 34, 495–500. https://doi.org/10.18280/ria.340415 (2020).

Article Google Scholar

Ponnusamy, V., Marur, D. R., Dhanaskodi, D. & Palaniappan, T. Deep learning-based X-ray baggage hazardous object detection—An FPGA implementation. Revue d’Intelligence Artificielle 35, 431–435. https://doi.org/10.18280/ria.350510 (2021).

Article Google Scholar

Wei, Y., Zhu, Z., Yu, H. & Zhang, W. An automated detection model of threat objects for X-ray baggage inspection based on depth wise separable convolution. J. Real-Time Image Proc. 18, 923–935. https://doi.org/10.1007/s11554-020-01051-1 (2021).

Article Google Scholar

Danso, S. et al. Security inspection image processing methods applying wavelet transform filters on terahertz active images. Revista de Investigaciones Universidad del Quindío 34, 37–51. https://doi.org/10.33975/riuq.vol34n1.853 (2022).

Article Google Scholar

Yasar, H. & Ceylan, M. A new deep learning pipeline to detect Covid-19 on chest X-ray images using local binary pattern, dual tree complex wavelet transform and convolutional neural networks. Appl. Intell. 51, 2740–2763. https://doi.org/10.1007/s10489-020-02019-1 (2021).

Article Google Scholar

Rahman, T. et al. Exploring the effect of image enhancement techniques on COVID-19 detection using chest X-ray images. Comput. Biol. Med. 132, 104319. https://doi.org/10.1016/j.compbiomed.2021.104319 (2021).

Article CAS PubMed PubMed Central Google Scholar

Rao, K., Bansal, M. & Kaur, G. Retinex-centered contrast enhancement method for histopathology images with weighted CLAHE. Arab. J. Sci. Eng. 47, 13781–13798. https://doi.org/10.1007/s13369-021-06421-w (2022).

Article CAS Google Scholar

Hättenschwiler, N., Mendes, M. & Schwaninger, A. Detecting bombs in X-ray images of hold baggage: 2D versus 3D imaging. Hum. Factors 61, 305–321. https://doi.org/10.1177/0018720818799215 (2019).

Article PubMed Google Scholar

Gao, Q., Hong, R., Zhu, X., Liu, X. An X-ray image enhancement algorithm for dangerous goods in airport security inspection. In Proceedings of Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS), Shenyang, China 43–46 (2021). https://doi.org/10.1109/ACCTCS52002.2021.00017.

Pincu, R., Kleinberger-Riedrich, O. A review digital radiography in the service of security. In 18th World Conference of Nondestructive Testing, 2012, 17, Durban, South Africa. www.ndt.net/?id=12810.

Gupta, P., Sinno, Z., Jack, L., Glover, J. L., Paulter, N. G. et al. Predicting detection performance on security X-ray images as a function of image quality. In IEEE Transactions on Image Processing, vol. 28, 3328–3342 (2019) https://doi.org/10.1109/TIP.2019.2896488.

Moldovanu, S., Damian Michis, F. A., Biswas, K. C., Culea-Florescu, A. & Moraru, L. Skin lesion classification based on surface fractal dimensions and statistical color cluster features using an ensemble of machine learning techniques. Cancers 13, 5256. https://doi.org/10.3390/cancers13215256 (2021).

Article PubMed PubMed Central Google Scholar

Montaha, S. et al. BreastNet18: A high accuracy fine-tuned VGG16 model evaluated using ablation study for diagnosing breast cancer from enhanced mammography images. Biology. 10, 1347. https://doi.org/10.3390/biology10121347 (2021).

Article PubMed PubMed Central Google Scholar

Montaha, S. et al. MNet-10: A robust shallow convolutional neural network model performing ablation study on medical images assessing the effectiveness of applying optimal data augmentation technique. Front. Med. 9, 924979. https://doi.org/10.3389/fmed.2022.924979 (2022).

Article Google Scholar

Verma, A. K. et al. Wavelet and deep learning-based detection of SARS-nCoV from thoracic X-ray images for rapid and efficient testing. Expert Syst. Appl. 185, 115650. https://doi.org/10.1016/j.eswa.2021.115650 (2021).

Article PubMed PubMed Central Google Scholar

Kusrini, K., Arif, M., Yudianto, A. & Al Fatta, H. The effect of Gaussian filter and data preprocessing on the classification of Punakawan puppet images with the convolutional neural network algorithm. Int. J. Electr. Comput. Eng. 12, 3752–3761. https://doi.org/10.11591/ijece.v12i4.pp3752-3761 (2022).

Article Google Scholar

Wu, Y., Dai, Q. & Lu, H. COVID-19 diagnosis utilizing wavelet-based contrastive learning with chest CT images. Chemom. Intell. Lab. Syst. 236, 104799. https://doi.org/10.1016/j.chemolab.2023.104799 (2023).

Article CAS Google Scholar

Hussein, F. et al. Hybrid CLAHE-CNN deep neural networks for classifying lung diseases from X-ray acquisitions. Electronics 11, 19. https://doi.org/10.3390/electronics11193075 (2022).

Article Google Scholar

High Tech Detection Systems, “SÛRETÉ & DÉTECTION,” (2023). https://www.htds.fr/ (accessed 27 Feb 2023).

Yu, L., Li, B. & Jiao, B. Research and implementation of CNN based on TensorFlow. IOP Conf. Ser. Mater. Sci. Eng. 490, 042022. https://doi.org/10.1088/1757-899X/490/4/042022 (2019).

Article Google Scholar

Jose, R. A convolutional neural network (CNN) approach to detect face using tensorflow and keras. Rochester, NY, 2019. https://papers.ssrn.com/abstract=3599641 (accessed 05 Mar 2023).

Download references

This work was supported by the project "DINAMIC", financed by the Romanian Ministry of Research and Innovation, Contract no. 12PFE/2021. The MSc student Lynda Oulhissane’s work was supported by the l’Agence Universitaire de la Francophonie (AUF), Programme “Eugen Ionescu” 2021–2022 (Bourse de recherche doctorale/postdoctorale). The author is extremely thankful for this.

Laboratory of Signals and Systems (LSS), Faculty of Science and Technology, Abdelhamid Ibn Badis University of Mostaganem, 11 Route Nationale, Kharouba, 27000, Mostaganem, Algeria

Lynda Oulhissane & Mostefa Merah

Department of Computer Science and Information Technology, Faculty of Automation, Computers, Electrical Engineering and Electronics, Dunărea de Jos University of Galati, 2 Stiintei Str., 800146, Galati, Romania

Simona Moldovanu

Modelling & Simulation Laboratory MSlab, Dunărea de Jos University of Galati, 47, 800008, Galati, Romania

Simona Moldovanu & Luminita Moraru

Department of Chemistry, Physics and Environment, Faculty of Sciences and Environment, Dunărea de Jos University of Galati, 47 Domneasca Str., 800008, Galati, Romania

Luminita Moraru

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

S.M., L.O., and L.M. designed research. L.O. and M.M. collected data. S.M., L.O., M.M. and L.M. analyzed data. L.O. and L.M. wrote the manuscript. All authors reviewed the manuscript.

Correspondence to Luminita Moraru.

The authors declare no competing interests.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and Permissions

Oulhissane, L., Merah, M., Moldovanu, S. et al. Enhanced detonators detection in X-ray baggage inspection by image manipulation and deep convolutional neural networks. Sci Rep 13, 14262 (2023). https://doi.org/10.1038/s41598-023-41651-y

Download citation

Received: 23 May 2023

Accepted: 29 August 2023

Published: 31 August 2023

DOI: https://doi.org/10.1038/s41598-023-41651-y

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.