X-ray security check image dangerous goods detection system based on improved YOLOv11

Abstract

The invention belongs to a target detection technology in the field of computer vision, and particularly relates to an X-ray security check image hazardous article detection system based on improved YOLOv11. Firstly, an X-ray security check image data set is divided into a training set and a test set according to the proportion of 6: 4, the system is based on a YOLOv11 architecture, a space and channel cooperative attention mechanism (SCSA) is introduced into a backbone network structure, the SCSA comprises two parts, namely sharable multi-semantic space attention (SMSA) and progressive channel self-attention (PCSA), and the SCSA is used as a main network of the system. The objective of the invention is to improve the detection performance of hazardous articles in an X-ray security inspection image, effectively enhance the understanding ability of a model for complex scenes, and improve the detection precision and speed of small targets, shielded objects and low-contrast hazardous articles. According to the method, the YOLOv11 algorithm is improved by combining the SCSA module, the detection precision and speed of the dangerous goods in the X-ray security check image are remarkably improved, and powerful technical support is provided for improving public safety and optimizing the security check process.

Description

Technical Field

[0001] The invention belongs to the technical field of target detection. Background Art

[0002] In recent years, the rapid development of deep learning technology has significantly changed the landscape of image processing and computer vision applications. In the field of security inspection, especially in transportation hubs such as airports, dangerous goods detection is of vital importance to ensuring public safety.

[0003] X-ray technology has become an important means of detecting high-density contraband hidden inside objects because it can penetrate a variety of substances and provide high-resolution images. Different materials have different absorption and scattering abilities for X-rays, which makes the X-ray images of goods present diverse color characteristics. Therefore, it is widely used in the field of security inspection. This technology has the advantages of less damage to goods and no need to unpack.

[0004] However, X-ray security inspections currently still rely mainly on manual monitoring. In traditional security inspection methods, security inspectors need to observe X-ray images with their naked eyes and judge whether there are dangerous goods based on experience. This results in the accuracy of detection being strongly dependent on the proficiency and mental state of the personnel, requiring high-intensity concentration for a long time, and prone to misjudgment due to fatigue. In addition, due to the low contrast, limited color range and overlapping objects of X-ray images, the detection process faces many difficulties. During peak security inspection periods, personnel fatigue and environmental factors may further increase the missed detection rate.

[0005] Therefore, using technologies such as deep learning to assist security inspection can effectively reduce human interference, improve detection accuracy, and process large amounts of data to increase detection speed. In order to improve the accuracy and efficiency of security inspections, automated and intelligent security inspection technologies are becoming a development trend. Especially in the field of X-ray image analysis, the application of computer vision technology has gradually become a key technology to improve security inspection efficiency and accuracy.

[0006] In the development of automated security inspection technology, deep learning algorithms have shown great potential in target detection. Among them, the YOLO (You Only Look Once) series of algorithms have attracted much attention in the field of computer vision due to their excellent real-time performance and accuracy. The YOLO architecture can simultaneously detect the location and category of multiple targets through a single forward propagation, greatly improving the detection efficiency. With the deepening of research, the YOLO algorithm continues to evolve, and each generation brings significant improvements. The latest YOLOv11 version has made breakthrough progress in detection accuracy and robustness, further improving its object detection capabilities in complex scenarios. These characteristics make the YOLO series of algorithms very suitable for use in fields such as security inspection that require high efficiency and high accuracy.

[0007] Although YOLOv11 performs well in many aspects, it still has certain limitations when facing special challenges in X-ray security images, such as detecting small or overlapping dangerous goods against complex backgrounds. This prompted us to think about how to further optimize the algorithm to better adapt to the specific needs of the security inspection field. Summary of the invention

[0008] The present invention aims to solve the problem of insufficient detection capability of small targets in complex backgrounds in X-ray security inspection images, and now provides an X-ray security inspection image dangerous goods detection system based on improved YOLOv11.

[0009] The present invention provides an X-ray security inspection image dangerous goods detection system based on improved YOLOv11, comprising:

[0010] Step 1. The present invention uses the PIDray dataset as the experimental data source. The dataset contains a large number of X-ray security images containing dangerous goods. In order to carry out effective model training and evaluation, we divide the entire dataset into two parts: a training set and a test set. The training set accounts for 60% of the total data volume and contains 29,457 images, which are used for model learning and parameter optimization. The test set contains the remaining samples, which is 40% of the total dataset, a total of 18,220 images, which are used to evaluate the actual performance and generalization ability of the model. This division ratio ensures sufficient training data while retaining enough test samples to verify the effectiveness of the model.

[0011] Step 2: Improve the YOLOv11 architecture and introduce the Spatial and Channel Synergistic Attention (SCSA) mechanism into the backbone network structure. It consists of two parts: Shareable Multi-Semantic Spatial Attention (SMSA) and Progressive Channel-wise Self-Attention (PCSA). It aims to improve the performance of dangerous goods detection in X-ray security inspection images, effectively enhance the model's ability to understand complex scenes, and improve the detection accuracy of small targets, occluded objects and low-contrast dangerous goods.

[0012] Step 3: Input the training data set into the improved YOLOv11 network (YOLOv11-SCSA) to carry out model training to build a deep learning model suitable for dangerous goods detection in X-ray security inspection images;

[0013] Step 4: Use the test data set to verify the performance of the trained YOLOv11-SCSA model, and comprehensively evaluate the detection effect of the model through multiple indicators.

[0014] Furthermore, in the present invention, in step 1, the public data set PIDray used in the experiment contains a total of 47,677 X-ray security inspection images, covering 12 categories of prohibited items such as firearms, knives, wrenches, pliers, scissors, etc. The data set is divided according to a 6:4 ratio, of which 29,457 images constitute training samples and the remaining 18,220 images are used as test samples.

[0015] In view of the difficulty of detecting contraband, the test set is subdivided into three levels: a single contraband scenario is classified as a simple mode, multiple contraband co-occurrences are classified as a difficult mode, and cases where contraband is deliberately concealed are classified as a hidden mode.

[0016] Furthermore, in the present invention, in step 2, the YOLOv11 architecture is improved, and the spatial and channel collaborative attention mechanism (SCSA) is introduced into the backbone network structure, which includes two parts: shareable multi-semantic spatial attention (SMSA) and progressive channel self-attention (PCSA). SMSA enhances the focus on different semantic areas, helping the model to detect key areas containing dangerous goods in X-ray security images, especially when objects are occluded, overlapped or irregularly shaped. The detection effect is better. PCSA helps the model focus on key feature channels related to dangerous goods by gradually weighting each channel, thereby improving the accuracy of feature extraction, especially for the detection of small objects. The SCSA module aims to improve the performance of dangerous goods detection in X-ray security images, effectively enhance the model's ability to understand complex scenes, and improve the detection accuracy and speed of small targets, occluded objects and low-contrast dangerous goods. The specific steps are:

[0017] The SCSA module first spatially splits the input tensor, average-pools the input tensor along the height and width directions, and then splits it into multiple sub-features to extract multi-scale spatial information. Next, convolution operations of different scales are applied to these sub-features to extract local and global features. The input features are weighted by spatial attention calculation combined with multi-scale convolution results. Next, the module processes the features using downsampling operations and calculates channel attention, generating weighted features through the self-attention mechanism of queries, keys, and values. Finally, the weighted features of spatial attention and channel attention are multiplied to obtain the output of the module, which further enhances the expressiveness of the features and improves the performance of the model.

[0018] Furthermore, in the present invention, in step 3, the training data set is input into YOLOv11-SCSA to carry out model training to construct a deep learning model suitable for dangerous goods detection in X-ray security inspection images, and the method is as follows:

[0019] 1. Modify the configuration file: Update the configuration file (cfg) of YOLOv11 and adjust the number of categories (classes) in the yaml file corresponding to the data to the number of categories actually marked in the dataset.

[0020] 2. Set hyperparameters: Configure the hyperparameters of the network model, including the size of the input image during training, the number of samples in each batch (batchsize), the number of training rounds (epochs), and the learning rate (learning rate).

[0021] Furthermore, in the present invention, in step 4, in order to verify the actual effect of the YOLOv11-SCSA model on the detection of dangerous goods in X-ray security inspection images, the performance of the trained YOLOv11-SCSA model is verified using a test data set, and the detection effect of the model is comprehensively evaluated through multiple indicators. Specifically:

[0022] The test set is fed into the YOLOv11-SCSA model to perform the dangerous goods detection task. The model analyzes each test image and outputs information about the detected suspicious items, including item category, location, and detection confidence. We then compare the model's predictions with the actual annotations of the test set and calculate various performance metrics, such as precision, recall, and F1 score. This comprehensive test not only evaluates the overall performance of the model, but also reveals its strengths and weaknesses in detecting different types of dangerous goods.

[0023] By analyzing the test results, we can gain a deeper understanding of the actual application effect of the YOLOv11-SCSA model and provide strong technical support for subsequent optimization and deployment. This systematic verification method can help us objectively evaluate the performance of the YOLOv11-SCSA model and ensure its reliability and effectiveness in actual security inspection environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is the process of the method for detecting dangerous goods in X-ray security inspection images based on improved YOLOv11 in the method of the present invention.

[0025] Figure 2 It is a schematic diagram of the network structure of X-ray security image dangerous goods detection of improved YOLOv11 in the method of the present invention.

[0026] Figure 3 It is a schematic diagram of the principle of the SCSA module added in the method of the present invention.

[0027] Figure 4 It is a comparison chart of the average accuracy of YOLOv11-SCSA and YOLOv11 in the method of the present invention.

[0028] Figure 5 It is a training indicator result diagram of YOLOv11-SCSA in the method of the present invention on the training set.

[0029] Figure 6 It is the performance index curve of YOLOv11-SCSA in the method of the present invention on the training set

[0030] Figure 7 It is a data distribution diagram of the data set used in the method of the present invention.

[0031] Figure 8 This is a visualization of the detection results of the improved method YOLOv11-SCSA in the method of the present invention on the test set. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0033] The X-ray security inspection image dangerous goods detection system based on improved YOLOv11 in this specific implementation method is as follows: Figure 1 As shown, the YOLOv11-SCSA network structure is as follows Figure 2 As shown, the following steps are included:

[0034] Step a: Use the public PIDray dataset as the experimental data source. The dataset contains a large number of X-ray security images containing dangerous goods. In this specific implementation, the PIDray dataset contains a total of 47,677 X-ray security images, covering 12 categories of prohibited items such as firearms, knives, wrenches, pliers, and scissors. The dataset is divided in a ratio of 6:4, of which 29,457 images constitute training samples and the remaining 18,220 images are used as test samples.

[0035] In view of the difficulty of detecting contraband, the test set is subdivided into three levels: a single contraband scenario is classified as a simple mode, multiple contraband co-occurrences are classified as a difficult mode, and cases where contraband is deliberately concealed are classified as a hidden mode.

[0036] Table 1. Distribution of the number of categories in the PIDRay dataset

[0037]

[0038]

[0039] Step b: Improve the YOLOv11 architecture. YOLOv11 mainly consists of three parts: Backbone, Neck and Head. The Spatial and Channel Synergistic Attention (SCSA) mechanism is introduced into the backbone network structure. The SCSA structure is as follows: Figure 3 shown.

[0040] SCSA spatial and channel collaborative attention consists of two parts: Shareable Multi-Semantic Spatial Attention (SMSA) and Progressive Channel-wise Self-Attention (PCSA). SMSA improves the focus on different semantic areas, helping the model detect key areas containing dangerous goods in X-ray security images, especially when objects are occluded, overlapped or irregularly shaped. PCSA helps the model focus on key feature channels related to dangerous goods by gradually weighting each channel, thereby improving the accuracy of feature extraction, especially for the detection of small objects. The SCSA module is designed to improve the performance of dangerous goods detection in X-ray security images, effectively enhance the model's ability to understand complex scenes, and improve the detection accuracy and speed of small targets, occluded objects and low-contrast dangerous goods.

[0041] The implementation principle of the SMSA module is as follows: First, the average pooling operation is performed on the input feature map to calculate the mean in the height and width directions, thereby generating two new feature maps x h and x w , respectively representing the height and width information in space. Next, x h Divided into four parts: local features l xh and three global features g at different scales xhs (small-scale global features), g xhm (global features at the mesoscale) and g xhl (large-scale global features), each part contains group_chans (i.e. C / 4) channels. The same operation is also applied to x w, and get the corresponding features. Then, these split features are processed by using four deeply shared 1D convolution kernels (the convolution kernel sizes are 3, 5, 7, and 9 respectively) to extract information of different scales and enhance the model's ability to handle multi-scale problems. Next is the calculation of spatial attention. The results extracted from each part of the feature are merged by concatenation (Concat), and then the normalization layer (GroupN orm) is applied for normalization. The spatial attention weight x is calculated by using the Sigmoid function hattn (x hattn represents the attention weight of the feature map in the height direction after being processed by the attention mechanism) and x wattn (x wattn represents the attention weight of the feature map in the width direction after being processed by the attention mechanism). Finally, the model dynamically adjusts the degree of attention of different spatial positions in the feature map through these spatial attention weights, thereby improving the expressiveness of the features.

[0042] The implementation principle of the PCSA module is as follows: First, the input feature map x(B, C, H, W) is downsampled. The downsampling method can be average pooling, maximum pooling or recombination. The goal is to obtain a more abstract feature representation by reducing the spatial resolution. After downsampling, the feature map x becomes a smaller size x(B, C, H′, W′), and then the feature map is normalized to ensure the consistency of data distribution. Next, by using a 1×1 deep convolution layer, the query (query, q), key (key, k) and value (value, v) are generated, which are the key components for calculating the relationship between channels. When calculating the attention matrix, the query and key are used to obtain the preliminary attention value through the dot product operation, and adjusted by the scaling factor to prevent numerical overflow. Then, the Sigmoid function is used to further calculate the final attention weight, and the Dropout technique is applied in the process for regularization to prevent overfitting. Finally, the obtained attention weights are matrix multiplied with the value (v) to obtain weighted features. These weighted features are output as channel enhanced feature maps, which ultimately enhance the importance of each channel in the feature map, thereby improving the expressiveness and recognition ability of the model.

[0043] Step c: Input the training data set into YOLOv11-SCSA to carry out model training to build a deep learning model suitable for dangerous goods detection in X-ray security inspection images.

[0044] Step c1: Input the training data set into YOLOv11-SCSA to carry out model training. The method for building a deep learning model suitable for dangerous goods detection in X-ray security inspection images is as follows:

[0045] Modify the cfg file of YOLOv11 and change the classes in the yaml file of the data to the number of categories marked by the dataset.

[0046] The hyperparameters of the network model are set, including the image input size of 640×640 for training the dataset and testing the model performance, the batch size of 64 input data per batch, the learning rate of 0.01, and the number of training rounds of 300.

[0047] Step c2: The present invention uses widely recognized evaluation indicators in the field of object detection to measure model performance, including recall, precision, F1-score, and mean average precision mAP@0.5 (mean average precision, IoU threshold is 0.5). The calculation formula is as follows:

[0048]

[0049] Among them, TP represents the number of correctly identified positive samples, and TN represents the number of correctly classified negative samples. FP refers to the number of negative samples that are misclassified as positive samples, while FN refers to the number of positive samples that are misclassified as negative samples. By calculating the precision and recall, we can get the F1 score, which is a comprehensive performance indicator. TP represents the positive sample that is accurately identified, f TF represents the negative samples that are accurately identified, and f TN It refers to the negative samples that are incorrectly identified. The area formed by the precision-recall curve (PR curve for short) and the coordinate axis can be used to calculate the average precision (AP) of each target category.

[0050] Step c3, the present invention aims to improve the target detection capability of the YOLOv11 algorithm, so we choose mAP@0.5 as the main evaluation index. In order to verify the improvement effect, we conducted a comparative experiment on the PIDRay dataset between YOLOv11 and the YOLOv11-SCSA model of the present invention.

[0051] Experimental results show that YOLOv11-SCSA has significant advantages in detection accuracy. Figure 4 The difference in average precision between the two models is presented intuitively, clearly demonstrating the superior performance of this model. This comparison result strongly proves that the YOLOv11-SCSA model can effectively improve the target detection capability.

[0052] This performance comparison verifies the effectiveness of the YOLOv11-SCSA model on specific tasks, and provides some technical support for further optimization of the target detection algorithm. The experimental results show that the YOLOv11-SCSA model has brought about a certain degree of performance improvement, which may be of some inspiration to subsequent related research.

[0053] Step d: Use the test data set to verify the performance of the trained YOLOv11-SCSA model.

[0054] Step d1: Use the test data set to verify the performance of the trained YOLOv11-SCSA model, and comprehensively evaluate the detection effect of the model through multiple indicators.

[0055] Step d2 and Table 2 present the performance comparison results of the YOLOv11-SCSA model and several mainstream target detection algorithms (including YOLOv5n, YOLOv10n and YOLOv11). It can be seen from the data that the YOLOv11-SCSA model outperforms other comparison algorithms in overall detection performance. And in the detection task of twelve types of dangerous goods in the data set, the present invention has achieved the highest detection accuracy in most categories. The improvement of the detection accuracy of the YOLOv11-SCSA model in low-accuracy categories (such as guns and knives) is particularly critical. It can better learn the characteristics of these small objects or difficult-to-identify objects, thereby improving the detection accuracy of these categories.

[0056] Table 2 Comparison of detection accuracy of different detection methods on PIDRay dataset

[0057]

[0058]

[0059] Figure 7 The size distribution of the ground-truth annotation boxes in the training dataset is shown. It can be clearly observed that the dataset is mainly composed of small objects, which reflects the particularity and challenge of the task.

[0060] Table 3 Model detection results of improved backbone network

[0061]

[0062] Table 3 shows that while maintaining the same amount of computation (FLOPs), YOLOv11-SCSA improves mAP by 1.9 percentage points and FPS by 149 frames per second compared to YOLOv11, providing higher detection accuracy and faster detection speed, which means that YOLOv11-SCSA has significant improvements in both performance and efficiency.

[0063] Figure 8 The detection effect of the YOLOv11-SCSA model on the test set is presented. The results show that the present invention can effectively detect dangerous goods targets in luggage, which confirms the excellent performance of the YOLOv11-SCSA model in dealing with complex scenes dominated by small targets.

[0064] By comparison Figure 7 and Figure 8 , we can better understand the potential of the present invention in practical applications, especially in X-ray security inspection scenarios where small dangerous objects need to be accurately detected. These results not only verify the effectiveness of the present invention, but also provide strong technical support for further optimizing small target detection technology.

[0065] Although the present invention has been described through specific embodiments, it should be noted that these examples are only used to illustrate the basic principles and applications of the present invention. Therefore, these embodiments can be modified and changed according to actual needs, and different configurations can be designed, as long as they do not deviate from the core idea and scope of the present invention as defined in the appended claims. In addition, multiple dependent claims and features of the present invention can also be used in combination in different ways. At the same time, the features of one embodiment can also be used in combination with the features of other embodiments.

Claims

1. A dangerous goods detection system for X-ray security inspection images based on improved YOLOv11, characterized in that: The following steps are involved: Step 1. The present invention uses the PIDray dataset as the experimental data source. The dataset consists of a large number of X-ray security images containing dangerous goods. In order to carry out effective model training and evaluation, we divide the entire dataset into two parts: a training set and a test set. The training set accounts for 60% of the total data volume and contains 29,457 images, which are used for model learning and parameter optimization. The test set contains the remaining samples, which is 40% of the total dataset, a total of 18,220 images, which are used to evaluate the actual performance and generalization ability of the model. This division ratio ensures sufficient training data while retaining enough test samples to verify the effectiveness of the model. Step 2: Improve the YOLOv11 architecture and introduce the Spatial and Channel Synergistic Attention (SCSA) mechanism into the backbone network structure. It consists of two parts: Shareable Multi-Semantic Spatial Attention (SMSA) and Progressive Channel-wise Self-Attention (PCSA). It aims to improve the performance of dangerous goods detection in X-ray security inspection images, effectively enhance the model's ability to understand complex scenes, and improve the detection accuracy of small targets, occluded objects and low-contrast dangerous goods. Step 3: Input the training data set into the improved YOLOv11 network (YOLOv11-SCSA) to carry out model training to build a deep learning model suitable for dangerous goods detection in X-ray security inspection images; Step 4: Use the test data set to verify the performance of the trained YOLOv11-SCSA model, and comprehensively evaluate the detection effect of the model through multiple indicators.

2. A dangerous goods detection system for X-ray security inspection images based on improved YOLOv11, characterized in that: The specific operation of dividing the X-ray security inspection image data set in step 1 is: The PIDray dataset used in the experiment contains a total of 47,677 X-ray security inspection images, covering 12 categories of prohibited items such as firearms, knives, wrenches, pliers, scissors, etc. The dataset is divided in a 6:4 ratio, of which 29,457 images constitute training samples and the remaining 18,220 images are used as test samples. In view of the difficulty of detecting contraband, the test set is subdivided into three levels: a single contraband scenario is classified as a simple mode, multiple contraband co-occurrences are classified as a difficult mode, and cases where contraband is deliberately concealed are classified as a hidden mode.

3. A dangerous goods detection system for X-ray security inspection images based on improved YOLOv11, characterized in that: In step 2, the spatial and channel collaborative attention mechanism (SCSA) is introduced into the backbone network structure, which includes two parts: shareable multi-semantic spatial attention (SMSA) and progressive channel self-attention (PCSA). SMSA improves the attention to different semantic areas, helping the model detect key areas containing dangerous goods in X-ray security images, especially when objects are occluded, overlapped or irregularly shaped. The detection effect is better. PCSA helps the model focus on key feature channels related to dangerous goods by gradually weighting each channel, thereby improving the accuracy of feature extraction, especially for the detection of small objects. The SCSA module aims to improve the performance of dangerous goods detection in X-ray security images, effectively enhance the model's ability to understand complex scenes, and improve the detection accuracy and speed of small targets, occluded objects and low-contrast dangerous goods. Specifically: The SCSA module first spatially splits the input tensor, average-pools the input tensor along the height and width directions, and then splits it into multiple sub-features to extract multi-scale spatial information. Next, convolution operations of different scales are applied to these sub-features to extract local and global features. The input features are weighted by spatial attention calculation combined with multi-scale convolution results. Next, the module processes the features using downsampling operations and calculates channel attention, generating weighted features through the self-attention mechanism of queries, keys, and values. Finally, the weighted features of spatial attention and channel attention are multiplied to obtain the output of the module, which further enhances the expressiveness of the features and improves the performance of the model.

4. A dangerous goods detection system for X-ray security inspection images based on improved YOLOv11, characterized in that: In step 3, adjust the model configuration file and set the hyperparameters. The specific steps are as follows: The image input size is 640×640, the input data size of each batch is 64, the learning rate is 0.01, and the number of training rounds is 300. The present invention adopts evaluation indicators widely recognized in the field of object detection to measure model performance, including recall, precision, F1-score and mean average precision mAP@0.5 (mean average precision, IoU threshold is 0.5).

5. A dangerous goods detection system for X-ray security inspection images based on improved YOLOv11, characterized in that: In step 4, the test set is used to evaluate the trained model. The specific steps are as follows: The present invention applies the YOLOv11-SCSA algorithm and the currently widely used detection models such as YOLOv5n, YOLOv10n, and YOLOv11 to the PIDray dataset for dangerous goods detection. Experimental results show that the improved detection model of the present invention has achieved significant improvements in detection accuracy and speed compared with other mainstream methods.

Images