X-ray image analysis method based on combination of improved Fast-RCNN and self-developed channel concentration model

Abstract

The invention relates to the technical field of computer vision and deep learning, in particular to an improved Fast-RCNN combined self-research channel concentration model X-ray image analysis method, which comprises the following steps of 1, inputting an X-ray image; step 2, feature extraction and channel concentration adjustment; step 3, region generation and target detection; and 4, outputting and optimizing a result. According to the X-ray image analysis method based on the improved Fast-RCNN in combination with the self-research channel concentration model, the sensitivity of the model to important features can be enhanced through the self-research channel concentration model, and the detection capability of a low-contrast X-ray image is improved; through weight adjustment, noise and irrelevant information in the image are suppressed, and the detection accuracy is improved; through combination of the deep convolutional network and a self-research attention mechanism, image features can be efficiently extracted, and the overall detection performance is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and deep learning technology, and specifically to an X-ray image analysis method combining an improved Fast-RCNN with a self-developed channel concentration model. Background Art

[0002] X-ray images are mainly used in medical diagnosis, security inspection and other fields. They are characterized by white background grayscale images, containing a large amount of monochrome information and slight grayscale changes. In traditional target detection tasks, the Fast-RCNN algorithm is widely used for target detection and recognition. However, the low contrast and grayscale change characteristics of X-ray images limit the detection performance of Fast-RCNN on these images.

[0003] Existing improvement methods mainly focus on optimizing loss functions, improving the region proposal network (RPN), and the depth of the network structure, but these methods fail to fully consider the characteristics of the X-ray image itself. Therefore, a new method is urgently needed to enhance the model's sensitivity to X-ray images and its feature extraction capabilities. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the prior art, the present invention provides an X-ray image analysis method combining an improved Fast-RCNN with a self-developed channel attention model.

[0006] (II) Technical solution

[0007] To achieve the above object, the present invention provides the following technical solution: an X-ray image analysis method combining an improved Fast-RCNN with a self-developed channel focus model, comprising the following steps:

[0008] Step 1: normalize the original X-ray image, enhance the contrast, and convert it into a monochrome grayscale image;

[0009] Step 2: Input the preprocessed X-ray image into the improved Fast-RCNN model, which is divided into four parts: feature extraction network, region proposal network, RoI pooling layer and fully connected layer. The feature extraction network uses a deep convolutional network as the basis, and inserts a channel attention module after each convolutional layer.

[0010] Step 3: After each layer of feature extraction, the channel attention module is used to calculate the weights of each channel and adjust the feature map. The channel attention module includes global average pooling, double-layer attention weight calculation and weight adjustment steps;

[0011] Step 4: Generate candidate regions using RPN, and resize and aggregate features of the candidate regions through the RoI pooling layer;

[0012] Step 5: Output the classification results and bounding box regression results in the fully connected layer to identify the target object in the X-ray image;

[0013] Step 6: Apply non-maximum suppression and other post-processing algorithms to remove redundant detection boxes and improve detection accuracy.

[0014] Preferably, the global average pooling step of the channel attention module is: given an input feature map, where H, W, and C represent the height, width, and number of channels of the feature map, respectively, global average pooling is performed on each channel.

[0015] Preferably, the double-layer attention weight calculation steps of the channel focus module are as follows: the weight of each channel is calculated through a double-layer fully connected network (FCN), the global feature vector is first input into the first fully connected layer, the output size of the layer is c / r, where r is the compression factor, to reduce the number of parameters and computational complexity, and the output is processed by the ReLU activation function σ1; then the output result is input into the second fully connected layer, and the weight of each channel is output, and the weight is processed by the Sigmoid activation function σ2.

[0016] Preferably, the calculated channel weight vector is applied to each channel of the original feature map to obtain an adjusted feature map.

[0017] Preferably, a channel focus loss term is added based on the original Fast-RCNN loss function to ensure the rationality of channel weight adjustment.

[0018] Preferably, during the training process, the classification loss, regression loss and channel focus loss are simultaneously optimized through the back propagation algorithm, so that the model can better focus on the salient features in the X-ray image while suppressing irrelevant or noise information.

[0019] (III) Beneficial effects

[0020] Compared with the prior art, the present invention provides an X-ray image analysis method combining an improved Fast-RCNN with a self-developed channel focus model, which has the following beneficial effects:

[0021] This improved Fast-RCNN combines the X-ray image analysis method with the self-developed channel focus model. The self-developed channel focus module can enhance the model's sensitivity to important features and improve the detection capability of low-contrast X-ray images.

[0022] By adjusting the weights, the noise and irrelevant information in the image are suppressed, and the detection accuracy is improved;

[0023] By combining deep convolutional networks and self-developed attention mechanisms, image features can be efficiently extracted and overall detection performance can be improved. DETAILED DESCRIPTION

[0024] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0025] An X-ray image analysis method combining an improved Fast-RCNN with a self-developed channel focus model includes the following steps:

[0026] Step 1: normalize the original X-ray image, enhance the contrast, and convert it into a monochrome grayscale image;

[0027] Step 2: Input the preprocessed X-ray image into the improved Fast-RCNN model, which is divided into four parts: feature extraction network, region proposal network, RoI pooling layer and fully connected layer. The feature extraction network uses a deep convolutional network as the basis, and inserts a channel attention module after each convolutional layer.

[0028] Step 3: After each layer of feature extraction, the channel attention module is used to calculate the weights of each channel and adjust the feature map. The channel attention module includes global average pooling, double-layer attention weight calculation and weight adjustment steps;

[0029] Among them, global average pooling:

[0030] Given an input feature map Where H, W, and C represent the height, width, and number of channels of the feature map, respectively. First, global average pooling is performed on each channel to obtain the global feature representation of each channel:

[0031]

[0032] Where ZC is the global average vector of channel c, with size RC.

[0033] Double-layer attention weight calculation:

[0034] Then, the weight of each channel is calculated through a two-layer fully connected network. First, the global feature vector z = [z 1 , z 2 , ..., z C ] is input to the first fully connected layer, the output size of this layer is Where r is the compression factor (usually set to 8 or 16) to reduce the number of parameters and computational complexity:

[0035] s=σ 1 (W 1 z+b 1 ),

[0036] Among them, σ1 is the ReLU activation function.

[0037] Next, S is input into the second fully connected layer to output the weight of each channel:

[0038] w=σ 2 (W 2 s+b 2 ),

[0039] Among them, σ2 is the Sigmoid activation function, w=[w 1 , w 2 , ..., w C ] is the weight vector for each channel;

[0040] Weight adjustment:

[0041] Apply the calculated channel weight vector w\mathbf{w}w to each channel of the original feature map to obtain the adjusted feature map:

[0042] F′ c =w c ·F c , c=1,2,...,C

[0043] Among them, F'C is the adjusted feature map, Fc is the original feature map, ω c is the weight of channel c.

[0044] Improvement of loss function:

[0045] Based on the original Fast-RCNN loss function, a channel focus loss term is added to ensure the rationality of channel weight adjustment:

[0046]

[0047] Lcls classification loss (using cross entropy loss).

[0048] Lreg bounding box regression loss (using smooth L1 loss).

[0049] Latt channel attention loss, defined as the mean square error (MSE) between the channel weight and its corresponding feature importance, is used to measure the feature extraction effect of the model after channel weighting.

[0050]

[0051] ∥Fc∥1 is the L1 norm of the feature map Fc, which is used to measure the importance of channel ccc.

[0052] λ is a weight factor used to balance the impact of the three loss terms.

[0053] Step 4: Generate candidate regions using RPN, and resize and aggregate features of the candidate regions through the RoI pooling layer;

[0054] Step 5: Output the classification results and bounding box regression results in the fully connected layer to identify the target object in the X-ray image;

[0055] Step 6: Apply non-maximum suppression and other post-processing algorithms to remove redundant detection frames and improve detection accuracy;

[0056] During the training process, the classification loss, regression loss and channel attention loss are optimized simultaneously through the back-propagation algorithm, so that the model can better focus on the salient features in the X-ray images while suppressing irrelevant or noise information.

[0057] The optimization goal is:

[0058]

[0059] Among them, θ is all the trainable parameters of the model;

[0060] In airport security inspection applications, the present invention can be used to improve the detection accuracy of potential threat items in luggage X-ray images and significantly reduce the false alarm rate and missed alarm rate.

[0061] In summary:

[0062] Input X-ray image: The system receives X-ray images and standardizes them to ensure that the image quality is suitable for subsequent processing.

[0063] Feature extraction and channel focus adjustment: The image passes through the feature extraction network, extracts features layer by layer, and adjusts the feature map at each layer using the self-developed channel focus module.

[0064] Region generation and target detection: Generate regions that may contain targets and perform target recognition and positioning.

[0065] Result output and optimization: Generate the final test results and optimize the results using post-processing methods.

[0066] This paper proposes an X-ray image analysis method that combines an improved Fast-RCNN with a self-developed channel focus model. By adaptively weighting each channel, the model's sensitivity to features in X-ray images and its detection accuracy are significantly improved.

[0067] In order to explain in detail the possible application scenarios, technical principles, specific schemes that can be implemented, and the purposes and effects that can be achieved, the following is a detailed description in conjunction with the specific embodiments listed. The embodiments described herein are only used to more clearly illustrate the technical solutions of the present application, and are therefore only used as examples, and cannot be used to limit the scope of protection of the present application.

[0068] Reference to "embodiment" herein means that the specific features, structures or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present application. The term "embodiment" appearing in various places in the specification does not necessarily refer to the same embodiment, nor does it particularly limit its independence or association with other embodiments. In principle, in the present application, as long as there is no technical contradiction or conflict, the various technical features mentioned in the embodiments can be combined in any way to form a corresponding implementable technical solution.

[0069] Unless otherwise defined, the technical terms used in this document have the same meanings as those generally understood by those skilled in the art to which this application belongs; the use of relevant terms in this document is only for describing specific embodiments and is not intended to limit this application.

[0070] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An X-ray image analysis method combining an improved Fast-RCNN with a self-developed channel focus model, characterized by: The following steps are involved: Step 1: normalize the original X-ray image, enhance the contrast, and convert it into a monochrome grayscale image; Step 2: Input the preprocessed X-ray image into the improved Fast-RCNN model, which is divided into four parts: feature extraction network, region proposal network, RoI pooling layer and fully connected layer. The feature extraction network uses a deep convolutional network as the basis, and inserts a channel attention module after each convolutional layer. Step 3: After each layer of feature extraction, the channel attention module is used to calculate the weights of each channel and adjust the feature map. The channel attention module includes global average pooling, double-layer attention weight calculation and weight adjustment steps; Step 4: Generate candidate regions using RPN, and resize and aggregate features of the candidate regions through the RoI pooling layer; Step 5: Output the classification results and bounding box regression results in the fully connected layer to identify the target object in the X-ray image; Step 6: Apply non-maximum suppression and other post-processing algorithms to remove redundant detection boxes and improve detection accuracy.

2. According to claim 1, the X-ray image analysis method of the improved Fast-RCNN combined with the self-developed channel concentration model is characterized by: The global average pooling step of the channel attention module is: given an input feature map, where H, W, and C represent the height, width, and number of channels of the feature map, respectively, global average pooling is performed on each channel.

3. The X-ray image analysis method of the improved Fast-RCNN combined with the self-developed channel focus model according to claim 2, characterized in that: The double-layer attention weight calculation steps of the channel focus module are as follows: the weight of each channel is calculated through a double-layer fully connected network (FCN), the global feature vector is first input into the first fully connected layer, the output size of this layer is c / r, where r is the compression factor, to reduce the number of parameters and calculation complexity, and the output is processed by the ReLU activation function σ1; then the output result is input into the second fully connected layer, and the weight of each channel is output, which is processed by the Sigmoid activation function σ2.

4. The X-ray image analysis method of the improved Fast-RCNN combined with the self-developed channel focus model according to claim 3, characterized in that: The calculated channel weight vector is applied to each channel of the original feature map to obtain the adjusted feature map.

5. The X-ray image analysis method of the improved Fast-RCNN combined with the self-developed channel focus model according to claim 4, characterized in that: Based on the original Fast-RCNN loss function, a channel focus loss term is added to ensure the rationality of channel weight adjustment.

6. The X-ray image analysis method of the improved Fast-RCNN combined with the self-developed channel focus model according to claim 5, characterized in that: During the training process, the classification loss, regression loss and channel attention loss are optimized simultaneously through the back-propagation algorithm, so that the model can better focus on the salient features in the X-ray images while suppressing irrelevant or noise information.

Images