An infrared small target detection method of an asymmetric feature enhancement network

Abstract

The application provides an infrared small target detection method of an asymmetric feature enhancement network, relates to the technical field of computer image segmentation, and comprises the following steps: pre-processing an input infrared image to represent a fixed-size feature tensor; defining a mixed convolution residual module, extracting image basis and directional enhancement features through double-branch convolution; adopting an asymmetric padding strategy for a spin-leaf type convolution to complete convolution and aggregation; defining a feature weighted fusion set, realizing double-branch feature fusion through a fusion weight alpha; combining an activation function, an attention mechanism and a residual connection processing to fuse features; continuously performing an encoder feature extraction operation on the features to screen small target discriminative features; obtaining intermediate layer features through a pooling convolution, and sequentially sampling and splicing multi-scale features through a decoder; generating multi-scale prediction maps and unifying resolutions, combining a gating mechanism to weight and fuse to output a target mask; and strengthening expression of infrared small target shallow features, accurately fusing multi-scale features and suppressing background interference.

Description

Technical Field

[0001] This invention relates to the field of computer image segmentation technology, and in particular to an infrared small target detection method using an asymmetric feature enhancement network. Background Technology

[0002] Infrared small target detection, a key technology in computer vision, aims to accurately identify and locate tiny targets in infrared images. Leveraging the characteristics of infrared imaging—based on thermal radiation sensing and requiring no visible light assistance—it plays an irreplaceable role in defense applications such as military reconnaissance, battlefield target early warning, and border security monitoring, as well as in civilian security surveillance and industrial equipment thermal anomaly detection. Infrared detection technology can overcome the limitations of harsh environments such as low light, no light, and fog, maintaining stable target perception capabilities in various low-visibility scenarios. However, infrared small targets inherently suffer from drawbacks—extremely small pixel ratio, low target-background contrast, and weak effective signals. Furthermore, in practical applications, they often encounter interference from complex backgrounds such as cloud clutter, building thermal radiation, and natural landscape textures, making it easy for small target features to be obscured by background information. This makes infrared small target detection consistently face the technical challenges of difficult feature extraction and unreliable target localization accuracy.

[0003] Traditional infrared small target detection methods have always relied on manually designed features and prior knowledge assumptions. They are mainly divided into three types: filtering, human visual system-inspired, and low-rank sparse decomposition. Although these methods are simple to implement and have low computational cost, they are poorly adaptable to complex backgrounds and are easily affected by image noise and background clutter. Furthermore, the detection performance of the model is highly dependent on manually tuned parameter settings, and its generalization ability in different practical application scenarios is significantly limited.

[0004] In recent years, the rapid development of deep learning technology has propelled it to become the mainstream research direction for infrared small target detection. The core idea of ​​existing related technical solutions is mostly to enhance the model's feature representation capabilities by constructing deeper network layers and designing more complex network structures. Techniques such as multi-scale feature fusion, channel-space attention mechanisms, and Transformer global feature modeling are widely used. However, in practical applications, these methods have exposed many unavoidable technical problems: the multiple downsampling operations accompanying increased network depth continuously weaken the already weak small target information during layer-by-layer feature extraction, sometimes even completely obscuring the features; complex network structures often focus on modeling high-level semantic information, neglecting key details of small targets such as edges, bright spots, and local textures contained in shallow features, directly leading to a significant decrease in the model's feature response capability for small targets; and excessively deep encoding structures also introduce a large amount of redundant semantic information unrelated to small targets, further interfering with the model's capture of effective features of small targets and reducing the model's sensitivity to small target recognition.

[0005] It is evident that both traditional and deep learning methods for infrared small target detection have their own technical shortcomings. Developing a novel detection method that can efficiently enhance the feature representation of small targets, accurately suppress background interference, and reduce redundant semantic modeling, thereby improving the accuracy, stability, and scene generalization ability of small target detection, has become an urgent need in this field. Summary of the Invention

[0006] This invention proposes an infrared small target detection method using an asymmetric feature enhancement network, aiming to solve the problems in existing infrared small target detection methods where the shallow features of small targets are weakened and deep redundant semantic interference is severe due to the increase in network depth.

[0007] This invention provides a method for infrared small target detection using an asymmetric feature enhancement network, the method comprising the following steps:

[0008] Step S1: Convert the input infrared image into a feature tensor;

[0009] Step S2: Extract the basic local features and directional enhancement features of the feature tensor through a hybrid convolutional residual module, wherein the hybrid convolutional residual module includes a standard convolutional branch and a swirl-shaped convolutional branch;

[0010] Step S3: Apply an asymmetric filling strategy to the spiral convolution to fill pixels, and then concatenate and aggregate the multi-directional convolution results;

[0011] Step S4: The basic local features and the directional enhancement features are weighted and fused to obtain the fused features;

[0012] Step S5: Apply channel attention, spatial attention, and residual connections sequentially to the fused features to obtain the activated features;

[0013] Step S6: Input the activated features into the encoder for multi-stage feature extraction, wherein the shallow stage of the encoder uses the hybrid convolutional residual module, the deep stage uses the standard residual block, and the discriminative features output at each stage are recorded.

[0014] Step S7: The encoder output features are sent to the decoder, and the decoder upsamples them layer by layer and concatenates them with the features of the corresponding encoder stage.

[0015] Step S8: Generate multi-scale prediction maps from each layer of the decoder, stitch the prediction maps together after unifying the resolution, generate a gating weight map through a gating mechanism, multiply the stitched features element-wise with the gating weight map, and output the final target mask through convolution.

[0016] Furthermore, the specific method for using an asymmetric filling strategy to fill pixels in the spiral convolution in step S3, and for concatenating and aggregating the multi-directional convolution results, includes:

[0017] The filling rules of the asymmetric filling strategy are composed of quadruples. Defined as follows: (1) represents the number of fill pixels in the left, right, top, and bottom directions, respectively. The specific fill rules are as follows:

[0018]

[0019] Where k represents the kernel length of the spiral convolution; the first two branches use 1×k convolution, the last two branches use k×1 convolution, and after being spliced ​​in the channel dimension, they are then convolved and aggregated.

[0020] Furthermore, the specific method for weighted fusion of the basic local features and directional enhancement features in step S4 to obtain the fused features includes:

[0021] Given a feature tensor X, in the input infrared image I, the feature set of the feature tensor X after step S2 is fused into a feature vector set. It contains all the features of bi-branch convolutions starting from the feature tensor X, i.e.

[0022]

[0023] Where Fs is the basic local feature, F R For directional enhancement features; let α be the fusion weight, and the feature weighted fusion weight. As learnable parameters, they are adaptively optimized through backpropagation during model training to dynamically balance the feature contributions of ordinary convolutional branches and swirl-shaped convolutional branches.

[0024] The feature fusion result is as follows:

[0025] F=F s +αF R .

[0026] Furthermore, the specific method for applying channel attention, spatial attention, and residual connections sequentially to the fused features in step S5 to obtain the activated features includes:

[0027] Activation functions are used to compute the effect on each element in the feature fusion set, utilizing... This represents the calculation, that is, a feature vector. The output structure is:

[0028] ,

[0029] Among them, M c and M sIt refers to the channel attention weights and spatial attention weights of a given feature tensor, i.e., when F is renormalized to F2, M c and M s X is a weight matrix that matches the F2 dimension. r It is a residual mapping.

[0030] Furthermore, in step S6, the activated feature input is fed into the encoder for multi-stage feature extraction. The shallow stage of the encoder uses the hybrid convolutional residual module, and the deep stage uses standard residual blocks. The specific method for recording the discriminative features output at each stage includes:

[0031] Encoder No. The module selection rules for each stage are as follows:

[0032]

[0033] That is, only the first two shallow stages of the encoder use hybrid convolutional residual modules to enhance the ability to extract shallow detail features and respond to small infrared targets; the deep stages of the encoder use standard residual blocks ResBlock and pooling operations to complete feature fusion and downsampling, simplifying the network structure and reducing the weakening of small target information by deep redundant semantic modeling; at the same time, the decoder maintains the complete hierarchical structure and feature reconstruction capability.

[0034] Furthermore, in step S7, the encoder output features are fed into the decoder, and the decoder performs upsampling layer by layer and concatenates the features with the corresponding encoder stage channels. The specific method for this includes:

[0035] Pooling and convolution operations are performed on the output feature map to obtain intermediate layer features. The image's spatial resolution is then restored via a decoder step, and feature information from different levels of the encoder is fused. The specific operations are as follows:

[0036] ,

[0037] ,

[0038] in, This indicates a bilinear upsampling operation. Indicates the decoder's first Layer operations, This indicates a channel splicing operation.

[0039] Furthermore, in step S8, multi-scale prediction maps are generated from each layer of the decoder. After unifying the resolution of each prediction map, they are stitched together. A gating weight map is generated through a gating mechanism. The final target mask is then output by multiplying the stitched features element-wise with the gating weight map and performing convolution. The specific method includes:

[0040] Each decoder generates a prediction map through a 1×1 convolution. Then, by upsampling to unify the resolution, all predicted sizes are aligned with... Consistently, the multi-scale prediction maps with aligned dimensions are stitched together. After 3×3 convolution, batch normalization, and ReLU activation, a gate weight G is generated through 1×1 convolution and Sigmoid activation. The fused features are then element-wise multiplied with the gate weight map to achieve adaptive weighting. Finally, the final target mask P is output through a 3×3 convolutional layer.

[0041]

[0042] in, This represents an element-wise multiplication operation, where G is a gated weight graph that matches the dimension of the fused feature. This represents a 3×3 convolution operation.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] This invention utilizes a feature extraction module to fuse standard convolutional branches and swirl-shaped convolutional branches, enhancing the shallow layer's local response to small infrared targets. It also combines an asymmetric coding strategy to reduce the weakening of small target information by deep redundant semantic modeling, achieving synergistic optimization of feature extraction and reconstruction. This significantly improves the accuracy and stability of small infrared target detection, and the model has strong generalization ability and is adaptable to various practical application scenarios.

[0045] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description

[0046] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0047] Figure 1 This is a flowchart of an infrared small target detection method using an asymmetric feature enhancement network provided by the present invention;

[0048] Figure 2 This is a schematic diagram of the hybrid convolutional residual module in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the asymmetric feature enhancement network structure in an embodiment of the present invention. Detailed Implementation

[0050] The exemplary embodiments disclosed in this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0051] This invention discloses an infrared small target detection method using an asymmetric feature enhancement network, such as... Figure 1 , Figure 2 and Figure 3 As shown, it includes the following steps:

[0052] 1) Let Indicates the input infrared image, using Let represent the feature tensor after the input image has been transformed, and For the corresponding OK Columns We construct a 3D feature tensor and treat the entire infrared image as a sequence of 3D feature tensors. This provides a regular data structure for subsequent convolution operations and lays the foundation for spatial-channel joint modeling of local feature extraction and attention mechanisms.

[0053] 2) Define the Hybrid Convolutional Residual Module (HPRB) as follows: Figure 2 As shown, the shallow local features of the input image in step 1) are captured, and the basic local features and directional enhancement features of the image are extracted by dual-branch convolution, respectively. Represents a sequence Standard convolution operations, utilizing Represents a sequence The spiral convolution operation is used. Combining these two methods ensures the stability of the basic features while enhancing the ability to extract directional features from small targets.

[0054] 3) Regarding step 2), An asymmetric padding strategy is used to represent the pixel padding constraints of the lobed convolution, utilizing...

[0055]

[0056] The padding rule for spiral convolution, i.e., quadruples These represent the number of fill pixels in the left, right, top, and bottom directions, respectively. This indicates the kernel length of the spinner convolution; the first two branches use... Convolution, the last two branches Convolution is performed, and the concatenation is then performed after splicing the channels.

[0057] 4) Define a feature weighted fusion set, given a feature tensor. When inputting infrared images In the middle, after step 2) The feature fusion set is a feature vector set. It contains all from The initial features of the bi-branch convolution, namely

[0058]

[0059] Among them, let The feature fusion result is as follows: (This is the result of the feature fusion process, where weights are used for fusion.) In this way, the model can adaptively adjust the weights of the two features during training: when the background is clean and the small target is small, the weights can be increased. Enhance directional features; when the background noise is complex, the model can automatically reduce... It relies more on robust standard convolutional branches. This is more flexible and theoretically sound than fixed-weight concatenation. Experiments show... It represents the optimal balance between responsiveness and interference resistance.

[0060] 5) Calculate the activation function for each element in the feature fusion set from step 4), and utilize... This represents the calculation, i.e., an eigenvector. The output structure is: ,in, and These are the channel attention weights and spatial attention weights for a given feature tensor, i.e. After restructuring hour, and To and Weight matrix for dimension matching It is a residual mapping.

[0061] 6) Continue to perform encoder feature extraction operation on the output of step 5), select and record the most discriminative small target features in each encoder.

[0062] 7) Perform pooling and convolution operations on the feature map output from step 6) to obtain intermediate layer features. The image's spatial resolution is then restored via a decoder step, and feature information from different levels of the encoder is fused. The specific operation is as follows:

[0063] ,

[0064] ,

[0065] in, This indicates a bilinear upsampling operation. Indicates the decoder's first Layer operations, This indicates the channel stitching operation. The shallow layers (the first two layers) use Hybrid Convolutional Residual Modules (HPRB) to fully capture shallow details such as edges and corners of small infrared targets; the deeper layers use standard residual blocks to avoid complex convolutions amplifying noise or diluting weak signals, thereby simplifying deep semantic modeling while preserving the sensitive features of small targets.

[0066] 8) Each decoder passes through Convolution generates prediction graphs Then, by upsampling to unify the resolution, all predicted sizes are aligned with... To ensure consistency, the multi-scale prediction map channels, after being aligned in size, are stitched together. After convolution, batch normalization, and ReLU activation, through Convolution and Sigmoid activation generate gated weights Adaptive weighting is achieved by multiplying the fused features element-wise with the gating weight map, ultimately through... The convolutional layer outputs the final target mask. :

[0067] ,

[0068] in, This represents element-wise multiplication. For gating weight graphs that match the fusion feature dimensions, express Convolution operation.

[0069] In this embodiment, as Figure 3 As shown, the overall framework consists of an input layer, an encoder, a decoder, and an output layer. The encoder and decoder achieve multi-scale feature fusion through skip connections; the input layer receives... The 3D infrared image feature tensor is used to generate two sets of basic local and directional enhancement data through the first two layers of the encoder's hybrid convolutional residual module. The encoder uses standard residual blocks and pooling operations to fuse and downsample features, filtering discriminative features for small targets and aggregating neighborhood features, while simultaneously reducing feature size and network training parameters. The encoder output features are fed into the decoder, where they undergo layer-by-layer bilinear upsampling and channel concatenation to restore spatial resolution. Through a collaborative design of asymmetric extraction by the encoder and multi-scale reconstruction by the decoder, features of small targets at different levels are gradually extracted and fused. The outputs of each layer of the decoder are then processed... Convolution generates multi-scale prediction maps, which are then upsampled to a uniform resolution and stitched together. These maps are then adaptively weighted and fused using a gating mechanism, and finally... The convolution outputs the target mask.

[0070] Optionally, the feature weighting fusion weight in step 4) As learnable parameters, adaptive optimization is performed via backpropagation during model training to dynamically balance the feature contributions of ordinary convolutional branches and spiral convolutional branches; targeting , , Experiments were conducted with 7 typical values. Among them, in The optimal expression effect is achieved by fully utilizing the characteristic responses of shallow, small targets while avoiding over-enhancing and introducing background interference.

[0071] Optionally, the asymmetric coding strategy described in step 6) employs a differentiated structural design for the encoder and decoder, with the encoder's first... The module selection rules for each stage are as follows:

[0072]

[0073] That is, only the first two shallow stages of the encoder use hybrid convolutional residual modules to enhance the extraction and response capabilities of shallow detailed features of small infrared targets; the deep stages of the encoder use standard residual blocks ResBlock to simplify the network structure and reduce the weakening of small target information by deep redundant semantic modeling; at the same time, the decoder maintains the complete hierarchical structure and feature reconstruction capabilities.

[0074] Example

[0075] The model was compared on two public datasets: IRSTD-1k and NUDT-SIRST, two publicly available infrared small target datasets. IRSTD-1k contains 1001 infrared images with complex backgrounds, covering challenging backgrounds such as strong cluttered clouds, thermal radiation from building clusters, and mountains. The NUDT-SIRST dataset contains 1327 labeled infrared images and includes more diverse scenes, such as complex backgrounds like the sea, urban blocks, and woodlands. The IRSTD-1k dataset was divided into training and test sets in a 4:1 ratio, and the NUDT-SIRST dataset was divided into training and test sets in the same ratio. Specific statistics are shown in Table 1.

[0076] Table 1 Data Statistics

[0077]

[0078] The implementation details are as follows:

[0079] This method instance is implemented based on the PyTorch deep learning framework and was trained and tested on an NVIDIA GeForce RTX 4090 GPU. During the data preprocessing stage, the size of all input images was uniformly set to... During training, the Adam optimizer was used to optimize the network parameters, with an initial learning rate set to... The batch size is set to... The training model is This involves several epochs. During training, the training data is randomly shuffled and input into the network in batches for iterative optimization. The model learns on the training set and reports the final performance metrics on the test set. The results are shown in Table 2.

[0080] Table 2 Experimental Comparison

[0081]

[0082] As shown in Table 2, the proposed method outperforms most existing models on different datasets. It achieves state-of-the-art performance on the IRSTD-1k dataset, surpassing existing methods in IoU, Pd, and Fa metrics, with IoU reaching 68.85, Pd reaching 94.42, and Fa reaching 11.53. On the NUDT-SIRST dataset, the proposed method achieves state-of-the-art performance in IoU and Fa, while being near-optimal in Pd. Particularly noteworthy is its ability to maintain high detection capability even in complex scenes and under significant scale variations, demonstrating excellent representation of small targets and stable detection performance.

[0083] The parameters were tuned as follows:

[0084] Decoder-side network depth: An attempt was made to reduce the number of convolutional layers stacked in the decoder, referencing the shallow design of the encoder. Results show that when the number of decoder layers is less than the original design depth, the model's ability to restore the spatial resolution of small infrared targets significantly decreases with the reduction in the number of layers. Target boundary blurring and detail loss become more pronounced, leading to a continuous decrease in detection accuracy. Only when the decoder maintains the full depth of the original design can the effects of layer-by-layer upsampling and multi-scale feature fusion be fully utilized to accurately restore the image's spatial resolution and achieve effective reconstruction of small target features. Based on the experimental results, this invention ultimately adopts the full depth of the original decoder design, balancing feature reconstruction capability and detection accuracy with the model's computational efficiency and generalization performance.

[0085] Through the above verification experiments, an infrared small target detection method based on an asymmetric feature enhancement network was proposed, and the most discriminative small target features in the feature map were further selected and retained. Experiments were conducted on two datasets in this verification experiment. Compared with existing infrared small target detection methods, this method significantly improves accuracy, effectively suppresses background clutter interference, and exhibits superior detection performance and generalization ability in complex scenes.

[0086] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for infrared small target detection using an asymmetric feature enhancement network, characterized in that, The method includes the following steps: Step S1: Convert the input infrared image into a feature tensor; Step S2: Extract the basic local features and directional enhancement features of the feature tensor through a hybrid convolutional residual module, wherein the hybrid convolutional residual module includes a standard convolutional branch and a swirl-shaped convolutional branch; Step S3: Apply an asymmetric filling strategy to the spiral convolution to fill pixels, and then concatenate and aggregate the multi-directional convolution results; Step S4: The basic local features and the directional enhancement features are weighted and fused to obtain the fused features; Step S5: Apply channel attention, spatial attention, and residual connections sequentially to the fused features to obtain the activated features; Step S6: Input the activated features into the encoder for multi-stage feature extraction, wherein the shallow stage of the encoder uses the hybrid convolutional residual module, the deep stage uses the standard residual block, and the discriminative features output at each stage are recorded. Step S7: The encoder output features are sent to the decoder, and the decoder upsamples them layer by layer and concatenates them with the features of the corresponding encoder stage. Step S8: Generate multi-scale prediction maps from each layer of the decoder, stitch the prediction maps together after unifying the resolution, generate a gating weight map through a gating mechanism, multiply the stitched features element-wise with the gating weight map, and output the final target mask through convolution.

2. The infrared small target detection method using an asymmetric feature enhancement network according to claim 1, characterized in that, The specific method for applying an asymmetric filling strategy to the spiral convolution in step S3 and concatenating the multi-directional convolution results includes: The filling rules of the asymmetric filling strategy are composed of quadruples. Defined as follows: (1) represents the number of fill pixels in the left, right, top, and bottom directions, respectively. The specific fill rules are as follows: Where k represents the kernel length of the spiral convolution; the first two branches use 1×k convolution, the last two branches use k×1 convolution, and after being spliced ​​in the channel dimension, they are then convolved and aggregated.

3. The infrared small target detection method using an asymmetric feature enhancement network according to claim 1, characterized in that, The specific method for weighted fusion of the basic local features and directional enhancement features in step S4 to obtain the fused features includes: Given a feature tensor X, in the input infrared image I, the feature set of the feature tensor X after step S2 is fused into a feature vector set. It contains all the features of bi-branch convolutions starting from the feature tensor X, that is: Where Fs is the basic local feature, F R For directional enhancement features; let α be the fusion weight, and the feature weighted fusion weight. As learnable parameters, they are adaptively optimized through backpropagation during model training to dynamically balance the feature contributions of ordinary convolutional branches and swirl-shaped convolutional branches. The feature fusion result is as follows: F=F s +αF R 。 4. The infrared small target detection method using an asymmetric feature enhancement network according to claim 1, characterized in that, The specific method for applying channel attention, spatial attention, and residual connections sequentially to the fused features in step S5 to obtain the activated features includes: Activation functions are used to compute the effect on each element in the feature fusion set, utilizing... This represents the calculation, that is, a feature vector. The output structure is: ， Among them, M c and M s It refers to the channel attention weights and spatial attention weights of a given feature tensor, i.e., when F is renormalized to F2, M c and M s X is a weight matrix that matches the F2 dimension. r It is a residual mapping.

5. The infrared small target detection method using an asymmetric feature enhancement network according to claim 1, characterized in that, In step S6, the activated features are input into the encoder for multi-stage feature extraction. The shallow stage of the encoder uses the hybrid convolutional residual module, and the deep stage uses standard residual blocks. The specific method for recording the discriminative features output at each stage includes: Encoder No. The module selection rules for each stage are as follows: That is, only the first two shallow stages of the encoder use hybrid convolutional residual modules to enhance the ability to extract shallow detail features and respond to small infrared targets; the deep stages of the encoder use standard residual blocks ResBlock and pooling operations to complete feature fusion and downsampling, simplifying the network structure and reducing the weakening of small target information by deep redundant semantic modeling; at the same time, the decoder maintains the complete hierarchical structure and feature reconstruction capability.

6. The infrared small target detection method using an asymmetric feature enhancement network according to claim 1, characterized in that, In step S7, the encoder output features are fed into the decoder, and the decoder performs upsampling layer by layer and concatenates the features with the corresponding encoder stage channels. The specific method includes: Pooling and convolution operations are performed on the output feature map to obtain intermediate layer features. The spatial resolution of the image is then restored through the decoder step, and feature information from different levels of the encoder is fused. The specific operation is as follows: , , in, This indicates a bilinear upsampling operation. Indicates the decoder's first Layer operations, This indicates a channel splicing operation.

7. The infrared small target detection method using an asymmetric feature enhancement network according to claim 1, characterized in that, In step S8, multi-scale prediction maps are generated from each layer of the decoder. After unifying the resolution of each prediction map, they are stitched together. A gating weight map is generated through a gating mechanism. The final target mask is then output by multiplying the stitched features element-wise with the gating weight map and performing convolution. The specific method includes: Each decoder generates a prediction map through a 1×1 convolution. Then, by upsampling to unify the resolution, all predicted sizes are aligned with... Consistently, the multi-scale prediction maps with aligned dimensions are stitched together. After 3×3 convolution, batch normalization, and ReLU activation, a gate weight G is generated through 1×1 convolution and Sigmoid activation. The fused features are then element-wise multiplied with the gate weight map to achieve adaptive weighting. Finally, the final target mask P is output through a 3×3 convolutional layer. in, This represents an element-wise multiplication operation, where G is a gated weight graph that matches the dimension of the fused feature. This represents a 3×3 convolution operation.

Images