Image copy mobile forgery detection network based on depth homologous features

Abstract

The invention discloses a homologous matching duplicating, moving and forgery positioning method. The method comprises an input module, a feature extraction module, a homologous matching module and a decoding module. The method comprises the following steps: firstly, carrying out a series of pre-preprocessing on an image through an input module, setting training hyper-parameters of a model, and adjusting a proper image size; secondly, extracting bottom detail features of the image through a feature extraction module; then, four homologous features from shallow to deep are mined through a homologous matching module; and finally, the diversified homologous features are utilized to reconstruct a counterfeit mask so as to realize accurate copy mobile counterfeit positioning and source and target distinguishing. According to the method, an effective result is obtained on a plurality of data sets, and effective help is provided for solving the field of image copying mobile counterfeiting.

Description

Technical Field

[0001] The present invention relates to the field of copy-move forged images, and in particular to an image copy-move forgery detection network based on deep homology features. Background Art

[0002] The widespread use of image processing tools such as Photoshop, Meitu, and GIMP has made image manipulation incredibly simple, with processed images being virtually indistinguishable from the original. However, many tampering operators attempt to obscure the true meaning of the original image through a series of manipulations. This has led to a gradual decrease in the credibility of images when used as digital evidence, medical diagnosis, and social networking, resulting in serious negative impacts on the public.

[0003] A common form of image forgery is copy-and-move forgery. Copy-and-move forgery involves copying and moving a region of an image to a different location within the current image. Because the regions originate from the same image, consistent lighting can make detection more challenging. Copy-and-move forgery detection primarily involves two tasks: localizing copy-and-move forgeries and distinguishing between forged sources and targets. The former aims to locate forged sources and targets within the image, while the latter aims to distinguish between them.

[0004] The current mainstream methods for detecting image copy and move forgeries are divided into those based on manual features and those based on deep features. The former mainly relies on manually selected image texture features to detect image forgeries. This type of method requires the parameters of the features to be selected specifically based on the specific image. Therefore, although this type of method has good detection results, it has two key disadvantages: (1) Feature selection depends on manually selected parameters and cannot be applied to image forgeries in different scenarios; (2) Manually selected features cannot be applied to unseen forged images and have poor robustness. The latter relies on popular deep learning methods to extract the deep feature information of the image and then detect copy and move forgeries by comparing the similarity between the deep feature information. Due to the strong robustness of deep features, this type of method has achieved good detection results for forged images with different lighting, different angles, and different scales.

[0005] However, existing deep feature-based forgery detection methods use cosine similarity, Jaccard similarity, Euclidean distance, Manhattan distance, etc. to calculate the similarity between features, and locate the forged area based on the similarity between deep features. Since the forged area usually only occupies a small part of the image, only a small number of similarity features are effective for forgery detection. In addition, there are often a large number of highly similar backgrounds in the image, such as the sky, forests, lakes and seas, which will also interfere with the similarity of deep features. In summary, the feature similarity calculation method of existing deep forgery detection technology is too simple and lacks the mining of deep information.

[0006] To solve this problem, it is necessary not only to use the original similarity features to ensure the final decoding of the image, but also to select the deep features with the strongest correlation from the similarity features through various methods to accurately locate the forged area.

[0007] To this end, this paper proposes a deep homology feature-based image copy-move forgery detection network. This network uses a homology matching module to mine diverse homology features. Simultaneously, these diverse homology features are used to reconstruct forgery masks, enabling accurate copy-move forgery localization and source-target differentiation. Summary of the Invention

[0008] In order to solve the problem that the feature similarity calculation method of copy-move forgery detection technology is too simple and lacks the mining of deep information, the present invention proposes an image copy-move forgery detection network based on deep homology features, which is used to mine different forms of deep homology features. At the same time, these diverse homology features are used to reconstruct the forgery mask to achieve accurate copy-move forgery positioning. To achieve the above purpose, the technical solution adopted by the present invention is as follows: An image copy-move forgery detection network based on deep homology features, comprising the following steps:

[0009] S1 Image preprocessing: Before inputting into the network, the dataset is split, the image is enhanced, the image size is reset, and the image is normalized to reduce the computational scale.

[0010] S11 Image Enhancement: Images are transformed through random cropping, random contrast adjustment, random brightness adjustment, random saturation adjustment, random hue adjustment, Gaussian blur, random rotation, and grayscale transformation operations, making the model adaptable to more varied image types and improving model robustness.

[0011] S12 Resize image: Reset the image to the same size using some effective image resizing methods, including nearest neighbor interpolation, linear interpolation, and seam-carving.

[0012] S13 Image Normalization: By normalizing the image, the overall value distribution of the image is adjusted to a smaller range to reduce the computational scale of the neural network. Specific normalization methods include: Min-Max normalization and adaptive threshold normalization

[0013] The specific adaptive threshold normalization rules are as follows:

[0014] 1) Obtain the frequency statistics of each pixel by counting the pixel histogram in an image;

[0015] 2) Select the minimum pixel value and the maximum pixel value that appear most frequently as Min and Max;

[0016] 3) Perform the same calculation operation as Min-Max normalization

[0017] S14 Dataset Split: Split a public dataset into training, validation, and testing parts according to a certain ratio, and use other public datasets for testing.

[0018] S2 deep feature extraction module: uses three different structured feature extractors, namely mesh structure, sequence structure and graph structure, to extract deep features of the image.

[0019] S21 mesh structure: Converts the image into small image blocks and uses a mesh structure with a certain number of kernels as the main convolution to extract deep feature representations between image blocks.

[0020] S22 sequence structure: Convert the image into a series of sequence block representations, use self-attention as the main sequence structure, and extract deep feature representations between sequence blocks.

[0021] S23 graph structure: Convert image blocks into graph nodes with graph associations, use a graph structure based on point convolution and pixel position relationship as the main basis, and extract deep feature representations between graph nodes.

[0022] S3 homology matching module: performs different forms of homology feature calculations on the deep features extracted by S2 to obtain primary homology features F1, high-order homology features F2, local homology features F3 and global homology features F4.

[0023] S31 calculates the primary homology feature F1: the deep features extracted by S2 are matched with themselves to obtain the global homology component, and the homology between each feature is improved through a learnable softmax by suppressing the homology between themselves. By screening features with significant relationships, the basic homology representation in the features is obtained as the primary homology feature F1.

[0024] S32 calculates the high-order homologous feature F2: the primary homologous feature F1 obtained in S31 is subjected to channel excitation squeezing through the deep homologous mining module to aggregate the scattered homologous components and separate the information of only a small amount of homologous components to obtain the high-order homologous feature F2.

[0025] S33 calculates the local homology feature F3: The high-order homology feature F2 obtained in S32 is used through a pseudo decoding module to obtain a homology probability distribution feature map of pseudo copy move forgery prediction. HGM filters out the original strong homology traces and uses the remaining residual information distribution as a potential reference for copy move. At the same time, through Conv agg Part of the image depth features are filtered to construct supplementary homology information. Finally, HGM is fused with the filtered part of the image depth features to obtain the local homology feature F3.

[0026] S34 calculates the global homology feature F4: The global homology component of the local homology feature F3 obtained in S33 is further enhanced, and the global homology feature F4 is obtained by fusing it with the global homology component obtained in S31.

[0027] S4 decoding module: The F1, F2, F3, and F4 features obtained in S3 are fused by strong and weak homology arrangement, and the features are reconstructed using the full convolution decoding module. The mask map of the copy-move forged positioning or the mask map of the copy-move forged source and target is obtained through bilinear interpolation.

[0028] S41 Homology strength and weakness arrangement fusion: Since the strength of the homologies among the four features F1, F2, F3, and F4 are inconsistent, the features are spliced and fused in the order of strength from small to large to obtain F total

[0029] S42 full convolution decoding module: the F obtained in S41 total , a fully convolutional module is used for feature reconstruction.

[0030] The beneficial technical effects of the present invention are as follows: the present invention repeatedly mines the deep features of the image through a homology matching module to obtain four effective homology features. Compared with the features obtained by traditional similarity calculation, our method can mine more effective information and can also identify more similar background areas in the image, greatly improving the accuracy of copy, move and forgery positioning. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is an overall flow chart of an image copy and move forgery detection network based on deep homology features of the present invention;

[0032] Figure 2This is a network architecture diagram of an image copy and move forgery detection network based on deep homology features of the present invention;

[0033] Figure 3 This is a detailed structural diagram of a homology matching module of an image copy and move forgery detection network based on deep homology features of the present invention;

[0034] Figure 4 This is a detailed structural diagram of a deep homology mining module of an image copy and move forgery detection network based on deep homology features of the present invention;

[0035] Figure 5 This is a detailed structural diagram of a pseudo decoding module of an image copy and move forgery detection network based on deep homology features of the present invention;

[0036] Figure 6 This is a diagram of the homology feature screening process of an image copy and move forgery detection network based on deep homology features of the present invention;

[0037] Figure 7 This is an example of the copy-move forgery positioning result of an image copy-move forgery detection network based on deep homology features of the present invention;

[0038] Figure 8 This is an example of the source and target differentiation results of an image copy and move forgery detection network based on deep homology features of the present invention; DETAILED DESCRIPTION

[0039] The patent of this invention is mainly intended to solve the problem of locating copy-move forgeries. Since many forged samples contain many similar areas, existing methods cannot truly identify whether these areas are truly forged areas.

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the embodiments, but the scope of protection claimed in the present invention is not limited to the following specific embodiments.

[0041] like Figure 1-7 As shown, a deep homology feature-based image copy and move forgery detection network includes the following steps:

[0042] S1 Image preprocessing: Before inputting into the network, the dataset is split, the image is enhanced, the image size is reset, and the image is normalized to reduce the computational scale.

[0043] In a specific implementation scheme, the S1 image preprocessing comprises the following steps:

[0044] S11 Image Enhancement: Images are transformed through random cropping, random contrast adjustment, random brightness adjustment, random saturation adjustment, random hue adjustment, Gaussian blur, random rotation, and grayscale transformation operations, making the model adaptable to more varied image types and improving model robustness.

[0045] S12 resets the image size: adjusts the image size to 256*256 using a linear interpolation algorithm.

[0046] S13 Image normalization: By using the normalization method, the overall numerical distribution of the image is adjusted and mapped to the range of [0,1] to reduce the computational scale of the neural network.

[0047] An example of adaptive threshold normalization is as follows:

[0048] 1) Obtain the frequency statistics of each pixel by counting the pixel histogram in an image;

[0049] 2) Select the minimum pixel value and the maximum pixel value that appear most frequently as Min and Max;

[0050] 3) Perform the same calculation operation as Min-Max normalization.

[0051] S14 Dataset Split: The 10,000-image USC-ISI dataset was split into 80,000 training images, 10,000 validation images, and 10,000 test images. The CASIA v2.0, DEFACTO, Coverage, and COMOFOD datasets were also used as tests to evaluate the generalization performance of the model.

[0052] S2 deep feature extraction module: uses three different structured feature extractors, namely mesh structure, sequence structure and graph structure, to extract deep features of the image.

[0053] In a specific implementation scheme, the S2 deep feature extraction module includes the following steps:

[0054] S21 mesh structure: Converts the image into small image blocks and uses a mesh structure with a certain number of kernels as the main convolution to extract deep feature representations between image blocks.

[0055] Specifically, given a 3*256*256 image, efficientNet-b7 is used to extract the features of the 5th, 6th, and 7th layers, which are 80*32*32, 160*16*16, and 224*16*16, respectively. Bilinear interpolation is used to convert them into 80*40*40, 160*40*40, and 224*40*40. At the same time, they are spliced by channel to obtain 464*40*40 multi-scale texture features. Next, an ASPP module with a void rate of (6, 12, 24) is used to deep sample the multi-scale texture features to obtain 176*40*40 deep contour features. These are spliced by channel to 640*40*40, obtaining the deep feature representation extracted by the mesh structure.

[0056] S22 sequence structure: Convert the image into a series of sequence block representations, use self-attention as the main sequence structure, and extract deep feature representations between sequence blocks.

[0057] Specifically, VIT is used as the backbone network to extract the deep features of the sequential structure of the image.

[0058] S23 graph structure: Convert image blocks into graph nodes with graph associations, use a graph structure based on point convolution and pixel position relationship as the main basis, and extract deep feature representations between graph nodes.

[0059] Specifically, VIG is used as the backbone network to extract the deep features of the image graph structure.

[0060] S3 homology matching module: performs different forms of homology feature calculations on the deep features extracted by S2 to obtain primary homology features F1, high-order homology features F2, local homology features F3 and global homology features F4.

[0061] In a specific implementation scheme, the S3 homology matching module includes the following steps:

[0062] S31 calculates the primary homology feature F1: the deep features extracted by S2 are matched with themselves to obtain the global homology component, and the homology between each feature is improved through a learnable softmax by suppressing the homology between themselves. By screening features with significant relationships, the basic homology representation in the features is obtained as the primary homology feature F1.

[0063] Specifically, first, given a set of image depth features F∈R with a size of 512×16×16 512x16x16 , where 512 is the number of channels and 16×16 is the size of the feature image. For ease of understanding, F can be considered as 512 feature blocks of size 16×16.

[0064] (1) First normalize the feature block

[0065]

[0066] It is normalized yes The mean of yes The standard deviation of .

[0067] (2) Then calculate and The rough feature similarity p(i,j)

[0068] Specifically, given the feature and where i=(i r ,i c ),j=(j r ,j c ), the rough feature similarity p(i,j) is calculated as follows:

[0069]

[0070] in(·) T is the transpose operator.

[0071] For a given We are in all possible Repeat this process to form a score vector S X [i], that is:

[0072] S X [i]=[p(i,0),…,p(i,j),…,p(i,256)]

[0073] if The features of are matched, then the score S X [i][j](i≠j) should be compared with other ) scores. However, it cannot be ignored that S X The score of [i][j] (i=j) will be the most significant because it is a match between itself and this part of information will affect the subsequent homology information mining. Therefore, this part of information is suppressed by the following formula:

[0074]

[0075] Where i,j are equal to i′,j′, ⊙ represents the element-wise product, and σ represents the suppression rate, which is set to 0.05 by default. This operation results in a suppression matrix with the same dimensions as the score, suppressing the self-correlation of the diagonal part. In order to better highlight the significant difference between the score and other scores, the similarity metric score strategy between features is calculated using the following formula:

[0076]

[0077] Among them, softmax(·) i Indicates the overall probability of calculating the unsorted scores in the dimension of position i or j, α represents the learnable context control parameter, and ⊙ refers to the dot product. Here, we get an unsorted significant homology similarity measure ρ X [i][j]. The probability of the homology similarity metric already contains enough information as statistical information about potential copy-move forgery regions. However, too much information will have a negative effect on the subsequent feature reconstruction process. Regardless of the length L of the homology similarity metric, we select the top K scores with the largest probability values to form a new summary score vector, namely:

[0078] P X [k][j]=ρ X [k′][j]

[0079] Where k∈[0,K-1], k′ is the index with the highest probability ranking in dimension i of the original vector. By selecting the top K scores, we get the first primary homology feature F1, which is a tensor of shape K×w×h, namely:

[0080] F1=P X [k][j]

[0081] Specifically, the screening method is as follows Figure 6 As shown in Figure 2, significant feature information is collected directly in order. This method differs from autocorrelation calculation, which reorders the similarity scores. This approach causes the relative position of similarity information to change. Although subsequent convolution operations can recover from this problem, it also causes many targets to be lost.

[0082] S32 calculates the high-order homology feature F2: the primary homology feature F1 obtained in S31 is subjected to channel excitation squeezing by the deep homology mining module (DHMM), the scattered homology components are aggregated, and only a small amount of homology component information is separated to obtain the high-order homology feature F2, namely:

[0083] F2=DHMM(F1)

[0084] Specifically, the structure of the deep homology mining module (DHMM) and its channel excitation squeezing process are as follows: given a combined feature C*H*W, the feature is cleaned through a convolution combination block (Conv+BN+Relu), and then the feature is adaptively pooled to obtain C*1*1 to obtain the homology information proportion of each channel, and the channel information is collected through a 1-dimensional convolution block with a kernel number of 3. The channel information of the scattered homology components is aggregated, and the information of a small amount of homology components is separated to obtain a new channel homology component information ratio. Then, the dot product is returned to the cleaned combined feature. Finally, the feature is cleaned again through a convolution combination block (Conv+BN+Relu) to obtain the high-order homology feature F2.

[0085] S33 calculates the local homology feature F3: The high-order homology feature F2 obtained in S32 is passed through a pseudo decoding module (PDM) to obtain a homology probability distribution feature map (HGM) of pseudo copy move forgery prediction. HGM filters out the original strong homology traces and uses the remaining residual information distribution as a potential reference for copy move. At the same time, through Conv agg Part of the image depth features are filtered to construct supplementary homology information. Finally, HGM is fused with the filtered part of the image depth features to obtain the local homology feature F3, namely:

[0086] F3=PDM(F2)⊙Conv agg (F X )

[0087] Specifically, the pseudo decoding module (PDM) consists of two groups of convolution blocks (Conv+BN+Relu) and a group of classification blocks (Conv+Conv+Sigmoid). By convolutionally aggregating the high-order homology features F2, the homology probability distribution feature map (HGM) is obtained. agg There are two main methods:

[0088] 1) Use conventional convolution blocks to filter the deep feature information of 256 channels;

[0089] 2) Use the ASPP module with a void ratio of (12 24 36) to extract 256 more effective deep feature information;

[0090] Specific fusion method: Fusion by point multiplication, namely Conv agg Each channel of the selected image depth features is multiplied by HGM to enhance the potential representation of the homology between features. Since HGM is obtained by high-order homology features F2, F2 is obtained by F1, and F1 is obtained by global homology components ρ X After screening the significant relationship, HGM is obtained, so it is just ρ XPart of the information indicates that the local homology feature F3 is obtained here.

[0091] For example: suppose the homology probability distribution feature map (HGM) is in

[0092] 1) 1 is the homologous distribution point that HGM considers to be infinitely close to 1;

[0093] 2) 0 is a non-homologous distribution point that HGM considers to be infinitely close to 0;

[0094] Let Conv agg The depth features of some channels obtained are By multiplying each channel This allows filtering out non-homologous information in deep features.

[0095] S34 calculates the global homology feature F4: the local homology feature F3 of S33 and the global homology component statistics ρ of S31 X Perform matrix multiplication to obtain the global homology feature F4.

[0096] Further global homology component enhancement is performed, and the global homology feature F4 is obtained by fusing it with the global homology component obtained in S31, namely:

[0097] F4=F3softmax(ρ X [i][j]) i

[0098] Specific fusion method, ρ X The global homology component statistics symbolizing the depth features of the image are further enhanced by matrix multiplication with F3 to further enhance the potential representation of the homology between features, and the global homology feature F4 is obtained here.

[0099] For example: Let the global homology component be Perform softmax operation on the homology component column to obtain Then set F3 to in

[0100] 1) 1 is the homologous significant feature point considered by F3;

[0101] 2) 0 is a non-significant feature point considered by F3;

[0102] Obtained by matrix calculation The global homology component enables F3 to obtain significant feature points that it had not discovered.

[0103] S4 decoding module: The F1, F2, F3, and F4 features obtained in S3 are fused by strong and weak homology arrangement, and the features are reconstructed using the full convolution decoding module. The mask map of the copy-move forged positioning or the mask map of the copy-move forged source and target is obtained through bilinear interpolation.

[0104] In a specific implementation scheme, the S4 decoding module includes the following steps:

[0105] S41 Homology strength and weakness arrangement fusion: Since the strength of the homologies among the four features F1, F2, F3, and F4 are inconsistent, the features are spliced and fused in the order of strength from small to large to obtain F total .Right now:

[0106] F total =cat(F3, F4, F1, F2)

[0107] S42 full convolution decoding module: the F obtained in S41 total , a fully convolutional module is used for feature reconstruction.

[0108] Specifically: Full convolutional decoding consists of three groups (Conv+Bn+Relu+Dropout) and is finally output through a 1*1Conv. It is divided into two cases according to different tasks:

[0109] 1) The output channel of the final Conv is 1, and the output value is passed through the sigmoid activation function and trained using the binary cross entropy loss. After training, the positioning prediction mask is obtained by binarization;

[0110] 2) The final Conv has 3 output channels and is trained using categorical cross entropy loss. After training, it is first passed through the softmax activation function and then binarized to obtain a predicted mask image that distinguishes the background, source, and target.

[0111] We have conducted ablation experiments on the USC-ISI, CASIAv2.0, DEFACTO, Coverage, and COMOFOD datasets to verify the effectiveness of the S3 homology matching module described in this paper. The F1-score metric used in the experiments is uniform.

[0112] Specifically, the USC-ISC training set was used for training, and the USC-ISC, CASIAv2.0, DEFACTO, Coverage, and COMOFOD datasets were used for testing. Under the influence of homology features of varying strengths, the generalization performance of the method was evaluated on datasets with different distributions. A comprehensive evaluation determined that the combination of the four features had the best generalization performance on the three datasets.

[0113]

Claims

1. A deep homology feature-based image copy and move forgery detection network, characterized by: The following steps are involved: S1 Image preprocessing: Before inputting into the network, the dataset is split, the image is enhanced, the image size is reset, and the image is normalized to reduce the computational scale. S2 deep feature extraction module: uses three different structured feature extractors, namely mesh structure, sequence structure and graph structure, to extract deep features of the image. S3 homology matching module: performs different forms of homology feature calculations on the deep features extracted by S2 to obtain primary homology features F1, high-order homology features F2, local homology features F3 and global homology features F4. S4 decoding module: The F1, F2, F3, and F4 features obtained in S3 are fused by strong and weak homology arrangement, and the features are reconstructed using the full convolution decoding module. The mask map of the copy-move forged positioning or the mask map of the copy-move forged source and target is obtained through bilinear interpolation.

2. The image copy and move forgery detection network based on deep homology features according to claim 1, characterized in that: The S1 image preprocessing includes the following steps: S11 Image Enhancement: Images are transformed through random cropping, random contrast adjustment, random brightness adjustment, random saturation adjustment, random hue adjustment, Gaussian blur, random rotation, and grayscale transformation operations, making the model adaptable to more varied image types and improving model robustness. S12 Resize image: Reset the image to the same size using some effective image resizing methods, including nearest neighbor interpolation, linear interpolation, and seam-carving. S13 Image normalization: By normalizing the image, the overall numerical distribution of the image is adjusted and mapped to the range of [0, 1] to reduce the computational scale of the neural network. S14 Dataset Split: Split a public dataset into training, validation, and testing parts according to a certain ratio, and use other public datasets for testing.

3. The image copy and move forgery detection network based on deep homology features according to claim 1, characterized in that: The S2 feature extraction module includes the following steps: S21 mesh structure: Converts the image into small image blocks and uses a mesh structure with a certain number of kernels as the main convolution to extract deep feature representations between image blocks. S22 sequence structure: Convert the image into a series of sequence block representations, use self-attention as the main sequence structure, and extract deep feature representations between sequence blocks. S23 graph structure: Convert image blocks into graph nodes with graph associations, use a graph structure based on point convolution and pixel position relationship as the main basis, and extract deep feature representations between graph nodes.

4. The image copy-move forgery detection network based on deep homology features according to claim 1, characterized in that: The S3 homology matching module includes the following steps: S31 calculates the primary homology feature F1: the deep features extracted by S2 are matched with themselves to obtain the global homology component, and the homology between each feature is improved through a learnable softmax by suppressing the homology between themselves. By screening features with significant relationships, the basic homology representation in the features is obtained as the primary homology feature F1. S32 calculates the high-order homology feature F2: the primary homology feature F1 obtained in S31 is subjected to channel excitation squeezing through the deep homology mining module (DHMM), the scattered homology components are aggregated, and only a small amount of homology component information is separated to obtain the high-order homology feature F2. S33 calculates local homology features F3: The high-order homology features F2 obtained in S32 are processed through a pseudo decoding module (PDM) to generate a homology probability distribution feature map (HGM) for pseudo copy move forgery prediction. The HGM filters out the original strong homology traces and uses the remaining residual information distribution as a potential reference for copy move. At the same time, we filter some image depth features to construct supplementary homology information. The HGM is then fused with the selected partial image depth features to obtain the local homology features F3. S34 calculates the global homology feature F4: S34 calculates the global homology feature F4: the local homology feature F3 of S33 and the global homology component statistics ρ of S31 X Perform matrix multiplication to obtain the global homology feature F4.

5. The image copy and move forgery detection network based on deep homology features according to claim 1, characterized in that: The S4 decoding module includes the following steps: S41 Homology strength and weakness arrangement fusion: Since the strength of the homologies among the four features F1, F2, F3, and F4 are inconsistent, the features are spliced and fused in the order of strength from small to large to obtain F total S42 full convolution decoding module: the F obtained in S41 total , a fully convolutional module is used for feature reconstruction.

Images