A lightweight camouflaged target detection method based on deep learning

Abstract

The present invention discloses a lightweight camouflaged target detection method based on deep learning. The method first obtains a camouflaged object image dataset, performs preprocessing, and divides it into a training set and a test set. Secondly, the test set data is input into a primary feature extraction network for primary feature extraction. The primary features are then enhanced through a spectrum gating unit, a hybrid attention mechanism, a global information module, an edge perception module, and an edge guidance module. Finally, based on deep separable convolution, the enhanced primary features are decoded and the detection results are output. The present invention makes full use of the RGB features and frequency features of the image and combines boundary guidance to achieve the detection of camouflaged targets.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a lightweight camouflaged target detection method based on deep learning. Background Art

[0002] Camouflaged object detection is an essential component of nature exploration, wilderness survival, and military exercises. Using computers to quickly and efficiently distinguish objects that blend seamlessly into a complex environment is a hot topic in computer vision.

[0003] The research behind camouflaged object detection stems from the complexity and uncertainty of the real world. With the advancement of technology, camouflage techniques are constantly improving, allowing objects to better blend into the background or change their appearance to evade detection by visual systems. This not only poses safety risks to people's daily lives but also increases the difficulty of security protection. Therefore, developing efficient and accurate camouflaged object detection technology has become a pressing issue in the field of computer vision.

[0004] The significance of camouflaged object detection lies in its wide range of applications and significant practical value. In security monitoring, it can be used to identify dangerous or contraband items disguised as ordinary objects, improving security inspection efficiency. In military defense, it can be used to detect military facilities or equipment disguised as natural objects, enhancing battlefield perception. In target tracking, it can be used to accurately track camouflaged targets in complex environments, improving tracking accuracy. Furthermore, camouflaged object detection helps promote the development and innovation of computer vision technology, providing new ideas and methods for research in related fields.

[0005] However, compared to traditional camouflaged object detection, research on lightweight camouflaged object detection is currently quite scarce. In the field of salient object detection (SOD), lightweight approaches often utilize lighter backbone networks. However, in the field of camouflaged object detection (COD), due to the greater complexity and uncertainty of the target task, simply applying lightweight methods from SOD has not been very effective. Currently, the mainstream approaches in this area rely on object edge guidance and frequency domain transformation.

[0006] As a branch of salient object detection, while camouflaged object detection has achieved considerable success, traditional methods struggle to accurately identify the complete and precise outline of target structures in complex lighting and scenes. Furthermore, these methods rely on parameter-heavy basic networks, making them unsuitable for the real-time demands of mobile devices. Therefore, considering the combined advantages of lightweight networks and the use of attention mechanisms for even better detection, experiments with deep neural networks for lightweight camouflaged object detection hold great research value and application prospects. Summary of the Invention

[0007] In response to the shortcomings of existing technologies and the high computational and time costs of traditional detection networks, the present invention provides a lightweight camouflaged target detection method based on deep learning.

[0008] The method of the present invention mainly includes three stages: primary feature extraction stage, enhanced feature extraction stage, and decoder stage. Each stage specifically includes the following steps:

[0009] Step (1). Obtain a dataset of camouflaged object images, perform preprocessing, and divide it into a training set and a test set.

[0010] Step (2). Primary feature extraction stage: input the test set data into the primary feature extraction network to perform primary feature extraction. The specific method is as follows:

[0011] This paper uses the backbone extraction model of MobileNet V3-Small to construct a lightweight primary feature extraction network. The 11 Backbone layers of the backbone network are merged into five units based on the number of channels. Training data is input, and the shallow information of the first three units and the deep information of the last unit are selected to generate four feature layers (f1, f2, f3, and f4), corresponding to 16, 24, 40, and 96 channels, respectively. These layers are then fed into the subsequent enhanced feature extraction stage.

[0012] Step (3). In the enhanced feature extraction stage, the primary features are enhanced through the spectrum gating unit, hybrid attention mechanism, global information module, edge perception module, and edge guidance module. The specific method is as follows:

[0013] This stage includes five basic modules, namely Spectral Gating Network (SGN), Convolutional Block Attention Module (CBAM), Global Information Module (GIM), Edge-Aware Module (EAM), and Edge-guidance Feature Modules (EFM).

[0014] (1) The spectrum gating unit first receives the outputs f1 and f2 of the first two large units from the primary feature extraction stage and uses them as its own input. It is responsible for processing shallow data with richer detailed information and learning the frequency features therein. The internal structure formula of this module is as follows:

[0015]

[0016]

[0017] Among them, LN is layer normalization, f i represents the real domain image information from the backbone network, X f represents frequency domain image information, FFT represents fast Fourier transform, IFFT represents inverse fast Fourier transform, MLP represents multi-layer perceptron, W c Represents the learnable weight parameter. The output of the spectrum gating unit is f i p , i∈{1,2}.

[0018] (2) There are four hybrid attention mechanisms, each receiving the output from two spectral gating units. And the outputs f3 and f4 of the last two units in the primary feature extraction stage. This module uses channel attention and spatial attention in series to strengthen the representation of important features. Its internal structure formula is as follows:

[0019] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0020] M s (F)=σ(f 7×7 ([AvgPool(M c (F));MaxPool(M c (F))]))

[0021] Among them, F is the input feature (including f3,f4), AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, MLP represents the multi-layer perceptron, σ represents the sigmoid activation function, and [;] represents the splicing operation. The output of the hybrid attention mechanism is M s (F): Contains

[0022] (3) The global information module is located in the deepest layer and receives the output f5 from the last unit of the backbone network. The global information module is used here to capture deep-level semantic features and pass them to the subsequent edge perception module. Its internal structure formula is as follows:

[0023] f′ a0 =Conv 1×1 (f5)

[0024] f′ a1 =Conv 3×3 ,rates=1(f5)

[0025] f′a2 =Conv 3×3 ,rates=2(f5)

[0026] f′ a3 =Conv 3×3 ,rates=3(f5)

[0027] f′ a4 =Conv 3×3 ,rates=4(f5)

[0028] f a =Conv 1×1 (Cat(f′ a0 ,f′ a1 ,f′ a2 ,f′ a3 ,f′ a4 ))

[0029] f high =up ×2 (CBR(f5+f a ))

[0030] Among them, rates represents the expansion rate of the dilated convolution operation (for example, when rates = 2, it means that the distance between the convolution kernel elements becomes 2, and the receptive field is expanded), Cat represents the concat operation, up ×2 Indicates that the upsampling ratio is 2. CBR is a combination of convolution Conv, batch normalization BatchNorm, and ReLU activation function. The output of the global information module is f high .

[0031] (4) The edge perception module receives the output feature f from the global information module high The output f1 of the first unit of the backbone network is used as its own input to merge shallow and deep information for object edge recognition. The module first uses two 1×1 convolutional layers to change the number of channels and high Upsample; secondly, the processed f1 and f high Perform concat splicing operation (vector splicing); finally, two 3×3 convolution layers and one 1×1 convolution layer with activation function are used to obtain effective edge features. The output of the edge perception module is the edge prediction map f e .

[0032] (5) The edge guidance module receives the output f from the edge perception module e , and combines the outputs from the hybrid attention mechanism at all levels This module first uses element-wise multiplication, residual connections, and 3×3 convolutions to generate a fused feature map. This fused feature map then passes through global average pooling (GAP), one-dimensional convolution, and a sigmoid activation function, and is then element-wise multiplied with the unprocessed feature map. Finally, a 1×1 convolution is performed to obtain the final output. Its internal structure formula is as follows:

[0033]

[0034]

[0035] Where D represents downsampling, represents element-wise addition, represents element-wise multiplication, is a 1D convolution with a kernel size of 1, and σ is the Sigmoid activation function. The output of the edge-guided module is f i a ,i∈{1,2,3,4}.

[0036] Step (4). In the decoder stage, based on depthwise separable convolution, the enhanced primary features are decoded and the detection results are output. The specific method is as follows:

[0037] The decoder uses a depthwise separable convolution (DWConv) instead of the standard convolutional layers. The DWConv-based Context Aggregation Module (DCAM) uses cross-scale interactions to enhance feature extraction.

[0038] This stage contains three layers of context aggregation modules. For the DCAM module of layer i (i=1,2,3), its input is the EFM output f from layer i. i a , EFM output of layer i+1 Inside the DCAM, the input is concatenated and subjected to 1×1 depthwise separable convolution to obtain f m After that, f m Divide into 4 parts according to the number of channels The four branches are passed in for processing (cross-level addition and expanded depth-separable convolution), and we get Where DWC represents the depth-separable convolution operation. The internal structure formula of DCAM is as follows:

[0039]

[0040] in represents a 3×3 depth-wise separable convolution, F DWConv1×1represents a 1×1 depthwise separable convolution, Represents the residual connection. The output of DCAM is the final mask prediction image f i c ,i∈{1,2,3}.

[0041] The three mask prediction images at different levels are trained for loss supervision, and the output of the bottom layer is As the final output, the other two are used as the intermediate layer outputs of the model and are only used for loss supervision; after training, the test set is used for verification.

[0042] The beneficial effects of the present invention are as follows:

[0043] The main advantages of the method of the present invention lie in five aspects: lightweight backbone network, frequency domain mechanism, global information module, introduction of attention mechanism, and lightweight context aggregation module for decoding.

[0044] The method of the present invention first uses a lightweight backbone network for feature extraction, fundamentally reducing network parameters and computational complexity while minimizing the need to significantly reduce network performance. It also introduces a frequency domain mechanism, less commonly used in RGB domain image processing, into the network to enhance frequency domain feature learning for better identification of camouflaged objects. The introduced hybrid attention mechanism effectively strengthens the representation of important features and suppresses features irrelevant to the task. The introduction of a global information module compensates for the weak deep extraction capabilities of the lightweight backbone network, and a lightweight context aggregation module is used for efficient decoding.

[0045] The network model provided by the method of the present invention can fully utilize the RGB features and frequency features of the image and realize the detection of camouflaged targets in combination with boundary guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a framework diagram of the method of the present invention;

[0047] Figure 2 This is a result diagram of the method of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described below with reference to the accompanying drawings.

[0049] like Figure 1 As shown, the present invention provides a lightweight camouflaged target detection method based on deep learning, and the specific steps are as follows:

[0050] Step (1). Primary feature extraction, the specific method is as follows:

[0051] The datasets used in this experiment are CAMO, COD10K, CHAMELEON, and NC4K, totaling four categories. With the exception of the CHAMELEON dataset, all other datasets are publicly available datasets in the COD field. The CAMO camouflaged object dataset contains 1,250 images of camouflaged objects in eight categories, including an edge image training set, a ground-truth object image training set, and an Img original image training set. COD10K is a large-scale, challenging dataset containing 10,000 images in 78 categories, annotated with high-precision, hierarchical, and logical annotations. The CHAMELEON dataset contains 76 manually annotated ground-truth images, collected from the internet using the Google search engine using the keyword "camouflaged animals." NC4K supports the localization and ranking of camouflaged objects, performing classification based on image information ranking. It contains 4,121 images of camouflaged objects with local information. Parts of the original dataset were used for both training and testing.

[0052] The training dataset uses the CAMO dataset, which contains 1,250 images of camouflaged objects in eight categories, and the COD10K dataset, which contains 2,790 samples of five object categories, for a total of 4,040 samples.

[0053] The test dataset uses 250 random samples from CAMO, 2026 remaining random samples from COD10K, 4121 random samples from NC4K, and 76 random samples from CHAMELEON.

[0054] First, a lightweight object feature extraction network was built, referring to the backbone extraction scheme of the MobileNet V3-Small version. This backbone network has a total of 14 layers. The present invention selects 11 layers of the basic structure Bneck, classifies and merges the Bneck layers into 5 units based on the number of output channels. The training set data is input, and the shallow layer information (the first three units) and the deep layer information of the last unit are used to generate feature layers f1, f2, f3, and f4, corresponding to 16, 24, 40, and 96 channels, respectively. These feature layers are then fed into the subsequent enhanced feature extraction stage.

[0055] Step (2). Spectrum gating unit module, the specific method is as follows:

[0056] The Spectral Gating Network (SGN) is a module that processes image features by converting them from the time domain to the frequency domain based on frequency information and Fourier transforms. Experimental studies have shown that the SGN's advantage lies in its ability to effectively identify local frequency features in an image, but its disadvantage lies in its inability to efficiently identify global semantic information. Because deep networks excel at processing global, abstract features, while shallow networks excel at processing local features, the SGN is more suitable for deployment in the shallow layers of the backbone network, where initial feature extraction takes place.

[0057] In this network, the SGN module receives the outputs f1 and f2 from the first two units of the backbone network and is responsible for processing shallow data with richer detailed information and learning the frequency characteristics. The formula is as follows:

[0058]

[0059]

[0060] Among them, LN is layer normalization, f represents the real domain image information from the backbone network, X f represents frequency domain image information, FFT represents fast Fourier transform, IFFT represents inverse fast Fourier transform, MLP represents multi-layer perceptron, W c Represents the learnable weight parameter. The output of the spectrum gating unit is f i p , i∈{1,2}.

[0061] Step (3). Hybrid attention mechanism module, the specific method is as follows:

[0062] Compared to the original ResNet backbone network, the MobileNet V3 series performs worse in primary feature extraction. To improve primary feature extraction, a hybrid attention mechanism (Convolutional Block Attention Module, CBAM) is used at the four feature layers of the MobileNet backbone network. By sequentially applying channel attention and spatial attention, it strengthens the representation of important features and suppresses features irrelevant to the task, thereby improving network performance.

[0063] The four hybrid attention modules receive the outputs from the two spectral gating units respectively. And the outputs f3 and f4 of the last two layers of the primary feature extraction stage. The formula is as follows:

[0064] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0065] Ms (F)=σ(f 7×7 ([AvgPool(M c (F));MaxPool(M c (F))]))

[0066] Among them, F is the input feature (including ), AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, MLP represents the multi-layer perceptron, σ represents the sigmoid activation function, and [;] represents the splicing operation. The output of the hybrid attention mechanism is M s (F): Contains

[0067] Step (4). Global information module, the specific method is as follows:

[0068] Low-level feature maps contain a wealth of detailed information, while deep-level feature maps contain comprehensive semantic information. Because the MobileNet backbone network's feature processing is inferior to traditional deep CNN networks, to improve the network's processing performance for deep-level features, a Global Information Module (GIM) is used to capture the deep-level semantic features f5 and pass them to subsequent modules.

[0069] The GIM module is similar to the dilated spatial convolutional pooling pyramid structure in that it processes the input data stream in parallel and outputs a data stream with the same number of channels. The GIM module is located after the last layer of the backbone network and receives deep abstraction information. This module obtains discriminative semantic information through the following process:

[0070] f′ a0 =Conv 1×1 (f5)

[0071] f′ a1 =Conv 3×3 ,rates=1(f5)

[0072] f′ a2 =Conv 3×3 ,rates=2(f5)

[0073] f′ a3 =Conv 3×3 ,rates=3(f5)

[0074] f′ a4 =Conv 3×3 ,rates=4(f5)

[0075] f a =Conv1×1 (Cat(f′ a0 ,f′ a1 ,f′ a2 ,f′ a3 ,f′ a4 ))

[0076] f high =up ×2 (CBR(f5+f a ))

[0077] Among them, rates represents the expansion rate of the dilated convolution operation (for example, when rates = 2, it means that the distance between the convolution kernel elements becomes 2, and the receptive field is expanded), at represents the concat operation, up ×2 Indicates that the upsampling ratio is 2. CBR is a combination of Conv, BatchNorm, and ReLU. The output of the global information module is f high .

[0078] Step (5). Edge perception module, the specific method is as follows:

[0079] Edge-Aware Module (EAM) receives features from shallow layers f1 and deep layers f high , and merge to realize edge recognition. First, two 1×1 convolution layers are used to change the number of channels, and f high Upsample; secondly, the processed f1 and f high Perform concat operation (vector concatenation); finally, obtain effective edge features through two 3×3 convolution layers and one 1×1 convolution layer with activation function. The formula is as follows:

[0080] f′1=Conv 1×1 (f1)

[0081] f′ high =Up(Conv 1×1 (f high ))

[0082] f e =σ(Conv 1×1 (Conv 3×3 (Conv 3×3 (f′1+f′ high ))))

[0083] Among them, Up represents the upsampling operation, Conv 1×1 and Conv 3×3represents the normal convolution with kernel size of 1 and 3, + represents the concat operation, and σ represents the sigmoid activation function. The output of the edge perception module is the edge prediction map f e .

[0084] Step (6). Edge-guided feature module, the specific method is as follows:

[0085] Edge-guidance Feature Modules (EFMs) enhance the feature representation of object structure by incorporating boundary information. EFMs use a channel-attention mechanism for cross-channel interaction, mining key information between different channels.

[0086] This module receives the output f from the edge perception module. e , and combines the outputs from the hybrid attention mechanism at all levels To enhance the feature representation of object structure. First, a fused feature map is obtained using element-wise multiplication, residual connection, and 3×3 convolution. This fused feature map is then passed through global average pooling (GAP), 1D convolution, and Sigmoid activation function, and then element-wise multiplication is performed with the unprocessed feature map. Finally, a 1×1 convolution is performed to obtain the output of the EFM module. The formula is as follows:

[0087]

[0088]

[0089] Where D represents downsampling, represents element-wise addition, represents element-wise multiplication, is a 1D convolution with a kernel size of 1, and σ is the Sigmoid activation function. The output of the edge-guided module is f i a ,i∈{1,2,3,4}.

[0090] Step (7). The lightweight context aggregation module in the decoder stage is as follows:

[0091] To further reduce model parameters while maintaining network processing power, this experiment optimized the standard convolution used in the decoder, replacing it with a depthwise separable convolution (DWConv). The DWConv-based Context Aggregation Module (DCAM) uses cross-scale interactions to enhance feature extraction.

[0092] For the DCAM module at layer i, the two inputs are the EFM output f at layer i. ia , EFM output of layer i+1 The input is concatenated and subjected to 1×1 depth-separable convolution to obtain f m After that, f m Divide into 4 parts according to the number of channels The four branches are passed in for processing (cross-level addition and expanded depth-separable convolution), and we get Where DWC represents the depth-wise separable convolution operation.

[0093] The internal structure formula of DCAM is as follows:

[0094]

[0095] in represents a 3×3 depth-wise separable convolution, F DWConv1×1 represents a 1×1 depthwise separable convolution, Represents the residual connection. The output of DCAM is the final mask prediction image f i c ,i∈{1,2,3}.

[0096] The three mask prediction images at different levels are trained for loss supervision, and the output of the bottom layer is As the final output, the other two are used as the intermediate layer outputs of the model and are only used for loss supervision; after training, the test set is used for verification.

[0097] Step (8). Presentation of invention results. Figure 2 As shown in the figure, the model can effectively detect camouflaged objects in most environments. It can also effectively segment the target object when the object outline is similar to the background outline (such as the hedgehog in the first row among a bunch of cacti), or the background texture is relatively complex (such as the gecko in the second row, which is located in a fallen leaf environment with a complex texture structure; the raptor in the third row, with a complex background of green leaves and tree trunks), or in most cases, the object and the background are the same color (such as the yellow and black snow leopard in the fourth row, the yellow hare in the fifth row, the black and white owl in the sixth row, the green turtle in the seventh row, and the goat in the eighth row).

Claims

1. A lightweight camouflaged target detection method based on deep learning, characterized in that: The following steps are involved: Step 1. Obtain a dataset of camouflaged object images, preprocess them, and divide them into a training set and a test set. Step 2. Input the test set data into the primary feature extraction network to perform primary feature extraction; Step 3. Enhance the primary features through the spectrum gating unit, hybrid attention mechanism, global information module, edge perception module, and edge guidance module in sequence. The specific process is as follows; Step 3.

1. The spectrum gating unit receives f1 and f2 as its own input and learns the frequency characteristics. The formula is as follows: Among them, LN is layer normalization, X f represents frequency domain image information, FFT represents fast Fourier transform, IFFT represents inverse fast Fourier transform, MLP represents multi-layer perceptron, W c Represents the learnable weight parameter; the output of the spectrum gating unit is f i p , i∈{1,2}; Step 3.

2. There are four hybrid attention mechanisms, each receiving the output from two spectral gating units. And the outputs f3 and f4 of the last two units of primary feature extraction; this module strengthens feature representation by serially using channel attention and spatial attention, and its formula is as follows: M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F))) M s (F)=σ(f 7×7 ([AvgPool(M c (F));MaxPool(M c (F))])) Among them, F is the input feature, including f3,f4, AvgPool represents the average pooling operation, MaxPool represents the maximum pooling operation, σ represents the sigmoid activation function, [;] represents the splicing operation; the output result of the hybrid attention mechanism is M s (F): Contains Step 3.

3. The global information module receives the output f5 from the backbone network, and the formula is as follows: f′ a0 =Conv 1×1 (f5) f′ a1 =Conv 3×3 ,rates=1(f5) f′ a2 =Conv 3×3 ,rates=2(f5) f′ a3 =Conv 3×3 ,rates=3(f5) f′ a4 =Conv 3×3 ,rates=4(f5) f a =Conv 1×1 (Cat(f′ a0 ,f′ a1 ,f′ a2 ,f′ a3 ,f′ a4 )) f high =up ×2 (CBR(f5+f a )) Among them, rates represents the expansion rate of the dilated convolution operation, Cat represents the concat operation, up ×2 Indicates that the upsampling ratio is 2. CBR is a combination of convolution Conv, batch normalization BatchNorm, and ReLU activation function. The output of the global information module is f high ; Step 3.

4. The edge perception module receives the output feature f from the global information module high The output f1 of the backbone network is used as its own input to perform object edge recognition. The module first uses two 1×1 convolutional layers to change the number of channels and high Upsample; secondly, the processed f1 and f high Perform concat operation; finally, two 3×3 convolution layers and one 1×1 convolution layer with activation function are used to obtain effective edge features, and the output result is the edge prediction map f e ; Step 3.

5. The edge guidance module receives the output f from the edge perception module e , and combines the outputs from the hybrid attention mechanism at all levels Enhance the feature representation of the object structure; this module first uses element-wise product, residual connection, and 3×3 convolution to obtain a fused feature map; then the fused feature map is passed through global average pooling GAP, one-dimensional convolution, and Sigmoid activation function, and element-wise product is performed with the unprocessed feature map, and finally the final output is obtained through 1×1 convolution. The formula is as follows: Where D represents downsampling, represents element-wise addition, represents element-wise multiplication, is a 1D convolution with a kernel size of 1, σ is the Sigmoid activation function, and the output of the edge guidance module is f i a ,i∈{1,2,3,4}; Step 4. Based on depth-wise separable convolution, decode the enhanced primary features and output the detection results.

2. The method for detecting lightweight camouflaged targets based on deep learning according to claim 1, characterized in that: The specific process of primary feature extraction is as follows: Based on the backbone extraction model of MobileNet V3-Small, a lightweight primary feature extraction network is constructed. The 11 Backkeck layers of the backbone network are merged into 5 units according to the number of channels. The training set data is input, and the first three units and the last unit are selected to generate four feature layers: f1, f2, f3, and f4.

3. The method for detecting lightweight camouflaged targets based on deep learning according to claim 2, characterized in that: The specific process of decoding the enhanced primary features is as follows: The lightweight context aggregation module based on the depthwise separable convolution DWConv uses cross-scale interaction to enhance the feature extraction capability. The module contains three layers of context aggregation modules. For the i-th layer, i = 1, 2, 3, the DCAM module, its input is the edge guidance module EFM output f from the i-th layer. i a , EFM output of layer i+1 Inside the DCAM, the input is concatenated and subjected to 1×1 depthwise separable convolution to obtain f m After that, f m Divide into 4 parts according to the number of channels Enter the four branches for processing respectively, and get Where DWC represents the depth-separable convolution operation; the internal structure formula of DCAM is as follows: in represents a 3×3 depth-wise separable convolution, F DWConv1×1 represents a 1×1 depthwise separable convolution, Represents residual connection; the output of DCAM is the mask prediction image f i c ,i∈{1,2,3}.

4. The method for detecting lightweight camouflaged targets based on deep learning according to claim 3, wherein: The step 4 also includes performing loss supervision training on the three mask prediction images of different levels, and As the final output, the other two are used as the intermediate layer outputs of the model and are only used for loss supervision; after training, the test set is used for verification.

Images