Zero watermark method and system based on improved lightweight neural network MobileNetV3

Abstract

The invention belongs to the technical field of medical image transmission, and particularly discloses a zero watermark method and system based on an improved lightweight neural network MobileNetV3, and the method comprises the following steps: S1, collecting an original medical image; s2, encrypting the original watermark image by using a Logistic logic chaotic system; s3, inputting the original medical image into an improved lightweight neural network MobileNetV3, and obtaining multi-scale depth features of the medical image; s4, based on the multi-scale depth features of the medical image, creating corresponding binary feature vectors by using a perceptual hash algorithm; and S5, performing XOR operation on the encrypted watermark image by using the binary feature vector to generate a zero watermark. By adopting the technical scheme, based on the improved lightweight neural network MobileNetV3, the method can focus on images of channel and spatial dimensions, the generalization ability and performance of the method are enhanced, it is ensured that the extracted medical features have excellent robustness, and the method is suitable for being applied to mobile medical care.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image transmission, and relates to a zero watermark method and system based on an improved lightweight neural network MobileNetV3. Background Art

[0002] In the information age, artificial intelligence technology has greatly improved production speed and work efficiency and has been widely used in many fields. With the popularization and promotion of advanced medical equipment and medical methods, original paper medical records and medical images have gradually been replaced by electronic medical records (EMR) and digital images.

[0003] While digital information provides a variety of convenient services, it also brings many information security issues. For example, information is prone to illegal tampering, copying, theft, leakage and other security issues during transmission. In addition, with the steady development of medical imaging technology and remote diagnosis, medical imaging is becoming more and more important as the basis for doctors' clinical diagnosis.

[0004] To facilitate diagnosis, doctors often embed personal information such as age, gender, and patient identification number in medical images. If the image is maliciously intercepted during network transmission, the patient's personal privacy information may be leaked, leading to problems such as copyright disputes over medical images and malicious theft of patient information. In addition, if the stored information is intercepted by criminals during the transmission of medical images and its content is tampered with, doctors may make incorrect diagnoses of patients' conditions, threatening their health. Therefore, effectively protecting the privacy of medical image information during the transmission of medical images is an urgent problem to be solved in the field of modern information security.

[0005] In order to solve the security problems based on digital information and protect the copyright information in medical images, digital watermarking technology has emerged. In 2003, Wenquan et al. first proposed the concept of zero watermark. The zero watermark algorithm uses the stable image features of the image itself to form authentication information, which has good robustness and imperceptibility. At present, many researchers have conducted extensive research on zero watermarks of medical images and proposed watermark algorithms that can solve the privacy and security issues of image information during transmission. However, most algorithms have problems such as complex watermark information embedding and extraction processes, low watermark embedding capacity, and still need to be improved in terms of geometric attacks.

[0006] With the advancement of artificial intelligence, deep learning has emerged as a cutting-edge technology and has gained great popularity. Since deep neural networks are good at extracting complex feature information, they are widely used in various fields such as natural language and computer vision. Therefore, some scholars combine deep learning technology with watermarking technology to improve the robustness of watermarking algorithms. Watermarking algorithms based on deep learning show good robustness in the face of common and geometric attacks. This method is very important for medical privacy protection. However, these network models have many parameters and high computing power requirements, which are not conducive to the practical application of mobile medical care. Summary of the invention

[0007] The purpose of the present invention is to provide a zero watermark method and system based on an improved lightweight neural network MobileNetV3, which reduces model parameters, computational complexity and space occupancy and is suitable for application in mobile health care.

[0008] In order to achieve the above object, the basic scheme of the present invention is: a zero watermark method based on an improved lightweight neural network MobileNetV3, comprising the following steps:

[0009] S1, collect original medical images;

[0010] S2, uses the Logistic logical chaos system to encrypt the original watermark image;

[0011] S3, uses the convolutional block attention module to replace the SE module in the lightweight neural network MobileNetV3, and adds a parallel depth-wise separable spatial pooling module after the second 1×1 convolution in the inverted residual block of MobileNetV3 to obtain an improved MobileNetV3;

[0012] Input the original medical image into the improved lightweight neural network MobileNetV3 to obtain the multi-scale deep features of the medical image;

[0013] S4, based on the multi-scale deep features of medical images, uses the perceptual hashing algorithm to create the corresponding binary feature vector;

[0014] S5, using the binary feature vector to perform an XOR operation on the encrypted watermark image to generate a zero watermark.

[0015] The working principle and beneficial effects of this basic solution are as follows: This technical solution integrates the convolutional block attention module (CBAM) into MobileNetV3, enabling the model to focus on images in channel and spatial dimensions, thereby enhancing its generalization ability and performance, and ensuring that the extracted medical features have excellent robustness. Adding parallel deep separable spatial pooling (PDSP) to the MobileNetV3 network can alleviate the gradient vanishing problem during model training and effectively extract multi-scale features of images.

[0016] Based on the improved lightweight neural network MobileNetV3, it has excellent robustness, reduces model parameters, computational complexity and space occupancy, and is very suitable for application in mobile healthcare.

[0017] Further, obtain the attacked medical image, execute steps S3 and S4, and obtain the corresponding binary feature vector F v ';

[0018] The binary feature vector F v 'Perform XOR operation with the generated zero watermark to obtain the watermark image W1':

[0019]

[0020] After performing logical chaos restoration on the extracted watermark image W1', the decrypted watermark image W' is obtained.

[0021] Simple operation and easy to use.

[0022] Furthermore, the original medical image is input into the improved lightweight neural network MobileNetV3 to obtain the multi-scale deep features F of the medical image. map (a,b,c), is:

[0023] F(i,j)→MobileNetV3→F map (a,b,c)

[0024] Among them, 1≤a≤8, 1≤b≤8, 1≤c≤512, F(i, j) represents the size of the original medical image, a, b represent the size of the multi-scale feature map output by the network, and c represents the number of channels of the feature map.

[0025] Effectively extract the multi-scale features of the image, which is beneficial for subsequent use.

[0026] Furthermore, based on the multi-scale deep features of medical images, the method of creating the corresponding binary feature vector using the perceptual hashing algorithm is as follows:

[0027] Fusion of multi-scale deep features of medical images map(a,b,c) to obtain the feature matrix F g (a,b):

[0028]

[0029] Use the perceptual hashing algorithm to generate a binary feature vector F v =:

[0030]

[0031] The algorithm is simple to operate and creates the corresponding binary feature vector.

[0032] Furthermore, the encrypted watermark image is XORed using the binary feature vector to generate a zero watermark Z(i,j), which is:

[0033]

[0034] Among them, i≤64, j≤64 represents the pixel value of the encrypted watermark image and the zero watermark key, W1 represents the encrypted watermark image, and K1 is the key.

[0035] The encrypted watermark image is operated using binary feature vectors to form a zero watermark, which can effectively resist common attacks and geometric attacks.

[0036] The present invention also provides a zero watermark system based on an improved lightweight neural network MobileNetV3, comprising an image acquisition module and a processing module, wherein the image acquisition module is used to acquire original medical images, and an output end of the image acquisition module is connected to an input end of the processing module;

[0037] The processing module executes the method of the present invention to generate and extract zero watermarks.

[0038] This system can capture the stable features of medical images across channels, spatial dimensions and multiple scales, has good resistance to common attacks and geometric attacks, and the watermark embedding and extraction process is efficient and convenient.

[0039] Furthermore, the processing unit is provided with an improved lightweight neural network MobileNetV3, a convolutional block attention module is used to replace the SE module in the lightweight neural network MobileNetV3, and a parallel depth-separable spatial pooling module is added after the second 1×1 convolution in the inverted residual block of MobileNetV3.

[0040] Integrating CBAM and PDSP modules into the MobileNetV3 architecture enables it to capture stable features of medical images across channels, spatial dimensions, and multi-scales, thereby facilitating the construction of a zero watermark algorithm.

[0041] Furthermore, the convolutional block attention module includes a channel attention module CAM and a spatial attention module SAM, and the feature matrix F∈R C×H×W The values ​​output from the deep convolutional layer in the inverted residual block of MobileNetV3 are used as the input of CBAM, where C represents the number of channels, and H and W represent the height and width of the corresponding feature map;

[0042] Perform global maximum pooling and global average pooling on the input feature matrix according to the channel to obtain two feature maps, which are then input into the multi-layer perceptron MLP;

[0043] The two feature maps generated by MLP are combined by pixel-by-pixel addition, and the channel weight vector M is generated using the Sigmoid activation function. c ;

[0044] The weight vector is multiplied by the input feature matrix to produce the channel attention feature map F';

[0045] Using the channel attention feature map F' as the input of SAM, by applying maximum pooling and average pooling along the spatial dimension, two different feature maps are obtained and concatenated to form a unified feature representation;

[0046] A 7×7 convolutional layer is used to compress the dimension to a single channel to obtain the spatial weight vector M s ;

[0047] M s Multiply it by F' to get the spatial attention feature map F".

[0048] Integrating the Convolutional Block Attention Module (CBAM) into MobileNetV3 enables the model to focus on images in both channel and spatial dimensions, thereby enhancing its generalization ability and performance and ensuring that the extracted medical features have excellent robustness.

[0049] Further, the dilated convolution in the atrous spatial pyramid pooling (ASPP) is replaced by a depthwise separable convolution to obtain the parallel depthwise separable spatial pooling module;

[0050] The depth-separable spatial pooling module includes convolutional layers arranged in parallel, a first depth convolutional layer with a rate of 2, a second depth convolutional layer with a rate of 4, a third depth convolutional layer with a rate of 6, and an image pooling layer;

[0051] The input of the depthwise separable spatial pooling module is connected to the output of the convolutional block attention module, and the convolution layer, the first depthwise convolution layer with rate 2, the second depthwise convolution layer with rate 4, the third depthwise convolution layer with rate 6, and the output of the image pooling layer are concatenated.

[0052] Adding Parallel Depthwise Separable Spatial Pooling (PDSP) to the MobileNetV3 network can alleviate the gradient vanishing problem during model training and effectively extract multi-scale features of images. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a structural schematic diagram of an improved inverted residual block of a lightweight neural network MobileNetV3 based on a zero watermark method of the present invention for improving the lightweight neural network MobileNetV3;

[0054] Figure 2 It is a structural schematic diagram of the convolution block attention module of the improved lightweight neural network MobileNetV3 of the zero watermark system based on the improved lightweight neural network MobileNetV3 of the present invention;

[0055] Figure 3 It is a structural schematic diagram of a deep separable spatial pooling module of an improved lightweight neural network MobileNetV3 of a zero watermark system based on an improved lightweight neural network MobileNetV3 of the present invention;

[0056] Figure 4 It is a flow chart of the zero watermark method based on the improved lightweight neural network MobileNetV3 of the present invention. DETAILED DESCRIPTION

[0057] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0058] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0059] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0060] The present invention discloses a zero watermark method based on an improved lightweight neural network MobileNetV3, which can effectively resist common attacks and geometric attacks, has strong robustness and low computational complexity, effectively protects the privacy of medical images, and brings great potential for the practical application of mobile medical health. Figure 4 As shown in the figure, the zero watermark method based on the improved lightweight neural network MobileNetV3 includes the following steps:

[0061] S1, collect original medical images;

[0062] S2, uses the Logistic logical chaos system to encrypt the original watermark image to improve its security;

[0063] S3, MobileNetV3 network adds squeeze and excitation lightweight attention module (SE) on the inverted residual block of MobileNetV2, and changes the Swish activation function to the H-Swish activation function, which has excellent performance and speed. However, the SE module is only calculated in the channel direction and cannot obtain attention in the spatial dimension.

[0064] Therefore, in order to improve the spatial feature extraction ability and accuracy of model training, the convolutional block attention module (CBAM) is used to replace the SE module in the lightweight neural network MobileNetV3, and a parallel depth-wise separable spatial pooling module (PDSP) is added after the second 1×1 convolution in the inverted residual block of MobileNetV3 to obtain an improved MobileNetV3, as shown in Figure 2. Figure 1 As shown;

[0065] Input the original medical image into the improved lightweight neural network MobileNetV3 to obtain the multi-scale deep features of the medical image;

[0066] S4, based on the multi-scale deep features of medical images, uses the perceptual hashing algorithm to create the corresponding binary feature vector;

[0067] S5, using the binary feature vector to perform an XOR operation on the encrypted watermark image to generate a zero watermark.

[0068] In a preferred embodiment of the present invention, the attacked medical image is obtained, and steps S3 and S4 are performed to obtain the corresponding binary feature vector F v ';

[0069] The binary feature vector F v 'Perform an XOR operation with the generated zero watermark to obtain the watermark image W1':

[0070]

[0071] After logical chaos restoration of the extracted watermark image W1', the decrypted watermark image W' is obtained. The robustness of the algorithm is measured by calculating the similarity between the original watermark image and the decrypted watermark image, where a higher similarity indicates that the algorithm has stronger robustness.

[0072] In a preferred embodiment of the present invention, the original medical image is input into the improved lightweight neural network MobileNetV3 to obtain the multi-scale depth feature F of the medical image. map (a,b,c), is:

[0073] F(i,j)→MobileNetV3→F map (a,b,c)

[0074] Among them, 1≤a≤8,1≤b≤8,1≤c≤512, F(i,j) represents the size of the original medical image, a, b represent the size of the multi-scale feature map output by the network, and c represents the number of channels of the feature map. For example, if we extract image features from the residual block of the tenth layer of the network, the size is F map (8,8,80).

[0075] In a preferred embodiment of the present invention, based on the multi-scale depth features of medical images, a method for creating a corresponding binary feature vector using a perceptual hashing algorithm is as follows:

[0076] Fusion of multi-scale deep features of medical images map (a,b,c) to obtain the feature matrix F g (a,b):

[0077]

[0078] Use the perceptual hashing algorithm to generate a binary feature vector F v =:

[0079]

[0080] In a preferred embodiment of the present invention, a binary feature vector is used to perform an XOR operation on the encrypted medical image to generate a zero watermark Z(i, j), which is:

[0081]

[0082] Where i≤64, j≤64, W1 represents the encrypted medical image, and K1 is the key. For security reasons, the zero watermark key Z and the key K1 generated by the logical scrambling system can be stored in a third-party copyright protection center.

[0083] The present invention also provides a zero watermark system based on an improved lightweight neural network MobileNetV3, comprising an image acquisition module and a processing module. The image acquisition module is used to acquire original medical images, and the output end of the image acquisition module is electrically connected to the input end of the processing module.

[0084] The processing module executes the method of the present invention to generate and extract zero watermarks. The system can capture stable features of medical images across channels, spatial dimensions and multiple scales, has good resistance to common attacks and geometric attacks, and the watermark embedding and extraction process is efficient and convenient.

[0085] In a preferred solution of the present invention, an improved lightweight neural network MobileNetV3 is provided in the processing unit, a convolutional block attention module is used to replace the SE module in the lightweight neural network MobileNetV3, and a parallel depth-separable spatial pooling module is added after the second 1×1 convolution in the inverted residual block of MobileNetV3.

[0086] Integrating CBAM and PDSP modules into the MobileNetV3 architecture enables it to capture stable features of medical images across channels, spatial dimensions, and multi-scales, thereby facilitating the construction of a zero watermark algorithm.

[0087] In a preferred embodiment of the present invention, Figure 2 As shown in Figure 2, the convolutional block attention module (CBAM) is a simple and effective attention component with rich representation capabilities and negligible parameter load. The module seamlessly integrates convolution and attention mechanisms. The convolutional block attention module includes a channel attention module CAM and a spatial attention module SAM. The feature matrix F∈R C×H×W The values ​​output from the deep convolutional layer in the inverted residual block of MobileNetV3 are used as the input of CBAM, where C represents the number of channels, and H and W represent the height and width of the corresponding feature map;

[0088] Perform global maximum pooling and global average pooling on the input feature matrix according to the channel to obtain two feature maps, which are then input into the multi-layer perceptron MLP;

[0089] The two feature maps generated by MLP are combined by pixel-by-pixel addition, and the channel weight vector M is generated using the Sigmoid activation function. c ;

[0090] The weight vector is multiplied by the input feature matrix to produce the channel attention feature map F';

[0091] Using the channel attention feature map F' as the input of SAM, by applying maximum pooling and average pooling along the spatial dimension, two different feature maps are obtained and concatenated to form a unified feature representation;

[0092] A 7×7 convolutional layer is used to compress the dimension to a single channel to obtain the spatial weight vector M s ;

[0093] M s Multiply it by F' to get the spatial attention feature map F".

[0094] Integrating the Convolutional Block Attention Module (CBAM) into MobileNetV3 enables the model to focus on the image in both channel and spatial dimensions, thereby enhancing its generalization ability and performance and ensuring that the extracted medical features have excellent robustness. The tandem connection of the two submodules in CBAM enables the network to process images from both spatial and channel dimensions, extract deeper texture features, and improve the generalization ability and performance of the model. CBAM gives the improved model better generalization ability and performance, enabling the construction of robust zero watermarks.

[0095] In a preferred embodiment of the present invention, Figure 3 As shown, the parallel depth-separable spatial pooling module is obtained by replacing the dilated convolution in the atrous spatial pyramid pooling ASPP with depth-separable convolution. The convolution kernels are divided into different groups using depth-separable convolution, which effectively expands the width of the network while compensating for any potential capacity loss. This method not only reduces the complexity and number of parameters, but is also a lightweight convolution method that requires fewer parameters and calculations than traditional convolution.

[0096] The depthwise separable spatial pooling module (PDSP) includes convolutional layers arranged in parallel, a first depthwise convolutional layer with a rate of 2, a second depthwise convolutional layer with a rate of 4, a third depthwise convolutional layer with a rate of 6, and an image pooling layer.

[0097] The input of the depthwise separable spatial pooling module is connected to the output of the convolutional block attention module, and the convolution layer, the first depthwise convolution layer with rate 2, the second depthwise convolution layer with rate 4, the third depthwise convolution layer with rate 6, and the output of the image pooling layer are concatenated.

[0098] Parallel Depthwise Separable Spatial Pooling (PDSP) is added to the MobileNetV3 network, using depthwise separable convolutions with different expansion rates to obtain features at multiple scales, and then aggregating multi-scale contexts to obtain global features, thereby enhancing the discriminability of feature extraction. This can alleviate the gradient vanishing problem during model training and effectively extract multi-scale features of images.

[0099] CBAM significantly improves the model's feature extraction capabilities, enhances its ability to resist common attacks, and lays the foundation for obtaining a robust watermark. The purpose of PDSP is to extract multi-scale features of images from the MobileNetV3 model to enhance resistance to geometric attacks, especially translation attacks. The combination of the two enables us to build a more robust zero-watermark method to resist common attacks and geometric attacks.

[0100] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0101] Although the 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 claims and their equivalents.

Claims

1. A zero watermark method based on an improved lightweight neural network MobileNetV3, characterized in that: The steps include: S1, collect original medical images; S2, uses the Logistic logical chaos system to encrypt the original watermark image; S3, uses the convolutional block attention module to replace the SE module in the lightweight neural network MobileNetV3, and adds a parallel depth-wise separable spatial pooling module after the second 1×1 convolution in the inverted residual block of MobileNetV3 to obtain an improved MobileNetV3; Input the original medical image into the improved lightweight neural network MobileNetV3 to obtain the multi-scale deep features of the medical image; S4, based on the multi-scale deep features of medical images, uses the perceptual hashing algorithm to create the corresponding binary feature vector; S5, using the binary feature vector to perform an XOR operation on the encrypted watermark image to generate a zero watermark.

2. The zero watermark method based on the improved lightweight neural network MobileNetV3 as claimed in claim 1, characterized in that: Get the attacked medical image, execute steps S3 and S4, and get the corresponding binary feature vector F v '; The binary feature vector F v 'Perform XOR operation with the generated zero watermark to obtain the watermark image W1': After performing logical chaos restoration on the extracted watermark image W1', the decrypted watermark image W' is obtained.

3. The zero watermark method based on the improved lightweight neural network MobileNetV3 as claimed in claim 1, characterized in that: The original medical image is input into the improved lightweight neural network MobileNetV3 to obtain the multi-scale deep features F of the medical image map (a,b,c), is: F(i,j)→MobileNetV3→F map (a,b,c) Among them, 1≤a≤8, 1≤b≤8, 1≤c≤512, F(i, j) represents the size of the original medical image, a, b represent the size of the multi-scale feature map output by the network, and c represents the number of channels of the feature map.

4. The zero watermark method based on the improved lightweight neural network MobileNetV3 as claimed in claim 1, characterized in that: Based on the multi-scale deep features of medical images, the method of creating the corresponding binary feature vector using the perceptual hashing algorithm is as follows: Fusion of multi-scale deep features of medical images map (a,b,c) to obtain the feature matrix F g (a,b): Use the perceptual hashing algorithm to generate a binary feature vector F v =:

5. The zero watermark method based on the improved lightweight neural network MobileNetV3 as claimed in claim 1, characterized in that: Use the binary feature vector to perform an XOR operation on the encrypted watermark image to generate a zero watermark Z(i,j), = W is 1(i:,j)Fv(k) Among them, i≤64, j≤64 represents the pixel value of the encrypted watermark image and the zero watermark key, W1 represents the encrypted watermark image, and K1 is the key.

6. A zero watermark system based on an improved lightweight neural network MobileNetV3, characterized in that: It includes an image acquisition module and a processing module, wherein the image acquisition module is used to acquire original medical images, and the output end of the image acquisition module is connected to the input end of the processing module; The processing module executes the method described in any one of claims 1-5 to generate and extract zero watermarks.

7. The zero watermark system based on the improved lightweight neural network MobileNetV3 as claimed in claim 6, characterized in that: The processing unit is provided with an improved lightweight neural network MobileNetV3, a convolutional block attention module is used to replace the SE module in the lightweight neural network MobileNetV3, and a parallel depth-separable spatial pooling module is added after the second 1×1 convolution in the inverted residual block of MobileNetV3.

8. The zero watermark system based on the improved lightweight neural network MobileNetV3 as claimed in claim 7, characterized in that: The convolutional block attention module includes a channel attention module CAM and a spatial attention module SAM, and the feature matrix F∈R C ×H×W The values ​​output from the deep convolutional layer in the inverted residual block of MobileNetV3 are used as the input of CBAM, where C represents the number of channels, and H and W represent the height and width of the corresponding feature map; Perform global maximum pooling and global average pooling on the input feature matrix according to the channel to obtain two feature maps, which are then input into the multi-layer perceptron MLP; The two feature maps generated by MLP are combined by pixel-by-pixel addition, and the channel weight vector M is generated using the Sigmoid activation function. c ; The weight vector is multiplied by the input feature matrix to produce the channel attention feature map F'; Using the channel attention feature map F' as the input of SAM, by applying maximum pooling and average pooling along the spatial dimension, two different feature maps are obtained and concatenated to form a unified feature representation; A 7×7 convolutional layer is used to compress the dimension to a single channel to obtain the spatial weight vector M s ; M s Multiply it by F' to get the spatial attention feature map F".

9. The zero watermark system based on the improved lightweight neural network MobileNetV3 as claimed in claim 7, characterized in that: Using depthwise separable convolution to replace the dilated convolution in the atrous spatial pyramid pooling ASPP to obtain the parallel depthwise separable spatial pooling module; The depth-separable spatial pooling module includes convolutional layers arranged in parallel, a first depth convolutional layer with a rate of 2, a second depth convolutional layer with a rate of 4, a third depth convolutional layer with a rate of 6, and an image pooling layer; The input of the depthwise separable spatial pooling module is connected to the output of the convolutional block attention module, and the convolution layer, the first depthwise convolution layer with rate 2, the second depthwise convolution layer with rate 4, the third depthwise convolution layer with rate 6, and the output of the image pooling layer are concatenated.

Images