A significant target recognition method based on Swin Transformer
By using the Swing Transformer and an improved loss function, the problems of insufficiently refined feature extraction and inadequate feature fusion in the backbone network were solved, resulting in a significant performance improvement in target recognition and reaching an advanced level.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF TECH
- Filing Date
- 2023-04-25
- Publication Date
- 2026-04-21
AI Technical Summary
Existing salient target recognition methods suffer from poor performance due to issues such as insufficient refinement of backbone network features, inadequate utilization of feature information during feature fusion, and fixed weights in the loss function.
Using the Swing Transformer as the backbone of the network architecture, and combining the channel calibration module, the cross-fusion module, and the PolyLoss redesigned cross-entropy loss function, a two-level decoder progressive feedback network is constructed to optimize the feature extraction and fusion process.
It improves the performance of salient target recognition, and can stably and efficiently extract and fuse feature information in various scenarios, reaching the state-of-the-art level to date.
Smart Images
Figure CN116597262B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, specifically to a salient target recognition method based on Swing Transformer. Background Technology
[0002] In recent years, deep learning has developed rapidly, and various models with significantly improved performance in computer vision have emerged one after another. Salient object recognition is an important research direction in computer vision, aiming to evaluate salient regions in images or videos, and is often used in the preprocessing part of visual tasks. Since Itti et al. first proposed the concept of salient object recognition in 1998, research on salient object recognition has gradually developed. Early salient object recognition methods mainly focused on how to find heuristic prior features, including color contrast, texture information, and center priors. However, these prior features are difficult to extract, usually requiring a lot of experience to extract good prior features. Furthermore, the fusion of the obtained prior features is difficult, making it hard to obtain the connection between semantic information and content. Although these heuristic methods can generate relatively coarse salient maps, their limitations in recognition accuracy and robustness make them difficult to apply in practical engineering projects. In recent years, the widespread application of convolutional neural networks in computer vision has demonstrated their powerful feature extraction capabilities. In particular, the introduction of VGG and ResNet has made it possible for deeper neural networks to achieve better performance. The successive proposals of numerous salient object detection models based on convolutional neural networks (CNNs) have propelled salient object recognition performance to new heights. These neural network-based salient object recognition models boast simple structures and excellent feature extraction capabilities. Since 2015, computer vision tasks have been dominated by CNNs. Starting with AlexNet and its revolutionary performance in the ImageNet image classification challenge, CNN architectures have evolved to become increasingly powerful through larger scales, wider connections, and more complex convolutional forms. As CNNs serve as the backbone networks for various vision tasks, these architectural advancements have led to performance improvements, thus broadly contributing to the flourishing of the entire field. On the other hand, the evolution of network architectures in Natural Language Processing (NLP) has taken a different path, with the Transformer becoming the most popular architecture today. Designed specifically for sequence modeling and transduction tasks, the Transformer is renowned for its use of attention to model long-term dependencies in data. Its tremendous success in the language domain has prompted researchers to investigate its adaptability to computer vision, and recently it has shown promising results on several tasks, particularly image classification and joint visual-language modeling. In this case, the Swin Transformer, a general visual backbone network, is used as the encoder layer to extend the applicability of the Transformer and use it as a general framework for salient object recognition to achieve the task of salient region segmentation.
[0003] In recent years, attention mechanisms have been widely applied to various network architectures, and numerous backbone networks with attention mechanisms have emerged. Since the advent of ResNet and VGG in 2015, deep learning researchers have made significant strides in various deep learning models using ResNet and VGG as backbone networks, driving the rapid development of deep learning. Meanwhile, the field of natural language processing has also made great strides under the guidance of Transformers. In 2019, Microsoft Research Asia overcame the challenges of the differences between natural language processing and computer vision, proposing a new hierarchical transformer architecture, Swin-Transformer. Swin Transformer can serve as a general backbone network for visual tasks, significantly breaking previous records in object detection and semantic segmentation tasks, and has been widely applied to numerous visual tasks, such as image generation, video action recognition, visual self-supervised learning, image restoration, and medical image segmentation. Swin Transformer has broken the long-standing dominance of CNNs (Convolutional Neural Networks) in the field of computer vision, accelerating the transformation of basic model architectures in computer vision. For a long time, computer vision tasks have been dominated by CNNs. The introduction of the Swin transformer has provided more inspiration for various vision tasks, and also provided more inspiration for the field of salient object recognition. The vigorous development of a field should be multifaceted, and CNNs, as a classic representative of capturing local features, are naturally regarded as the core of architectural design. However, with the progress of existing research, some classic CNNs have the following two main limitations:
[0004] First, CNNs use local receptive fields to capture features, so they cannot directly model the global environment;
[0005] Secondly, the weights learned by CNNs are static during inference, which makes CNNs unable to flexibly adapt to different input content.
[0006] Existing methods attempt to combine the superior features of the Swin Transformer with CNNs to address the aforementioned limitations.
[0007] Based on the Swin Transformer, a feature pyramid network based on Swin Transformer was designed, using it as a general backbone network.
[0008] Therefore, in order to meet the application needs of salient target recognition in computer vision, existing salient target recognition methods suffer from problems such as insufficient refinement of backbone network features, insufficient utilization of feature information during feature fusion, and fixed weights of loss functions, resulting in poor performance of salient target recognition. Therefore, a new solution is needed to address these issues. Summary of the Invention
[0009] The purpose of this invention is to provide a salient target recognition method based on Swing Transformer to solve the problems of poor performance in existing salient target recognition methods, such as insufficient refinement of backbone network features, insufficient utilization of feature information during feature fusion, and fixed weights of loss functions.
[0010] To achieve the above objectives, the present invention provides the following technical solution: a salient target recognition method based on Swing Transformer, comprising at least the following steps:
[0011] The Swing Transformer is used as the backbone of the network architecture to extract multi-level feature maps from images.
[0012] Establish a channel calibration module to recalibrate channel feature information and reduce the computational load of subsequent feature fusion;
[0013] A cross-fusion module is established to progressively fuse feature information by adding global information while multiplying and fusing. The cross-fusion module is a three-input, two-output module. The cross-fusion module multiplies and fuses the features of two adjacent layers, and then adds the sum of the corresponding high-level or low-level feature maps and the global feature map as a residual block to the corresponding high-level or low-level interface.
[0014] The loss function was redesigned using PolyLoss, and a Poly-based cross-entropy loss function was constructed to adjust the first term coefficient of the Taylor expansion of the BCE loss function in experiments to obtain better performance.
[0015] A two-stage decoder progressive feedback network is constructed using Swin Transformer as the backbone of the network architecture, a channel recalibration module, a cross-fusion module, and a Poly-based cross-entropy loss function to complete salient target recognition.
[0016] Preferably, the application of the channel calibration module includes at least the following steps:
[0017] Copy the feature information obtained from the backbone network;
[0018] The two sets of feature information are combined in parallel and fused.
[0019] By using a convolution to change the number of channels, the computational load of subsequent modules can be reduced.
[0020] Optimized channel feature information is obtained for subsequent feature fusion.
[0021] Preferably, the application of the cross-fusion module includes at least the following steps:
[0022] The high-level and low-level feature information optimized by the channel calibration module are sampled into images of the same resolution, and then the two feature maps are multiplied and fused to obtain a fused feature map.
[0023] The high-level or low-level feature maps are sampled to their respective resolutions. The high-level and low-level feature maps of the initial input are convolved, normalized, and activated, and then added as residual blocks to the fused features.
[0024] The global feature map is sampled to the resolution of both the high-level and low-level feature maps, and is also added as a residual block to the fused feature.
[0025] The high-level and low-level feature maps are respectively convolutionally normalized and activated to obtain new high-level and low-level feature maps for the next stage of feature fusion.
[0026] Preferably, the structural formula of the cross-fusion module is as follows:
[0027] p h '=Conv(Conv(Conv(Conv(p h ))*Cov(Conv(p l )))+Conv(p h )+Conv(p g ))
[0028] p l '=Conv(Conv(Conv(p h ))*Conv(Conv(p l )))+Conv(p l )+Conv(p g ))
[0029] Where, p h p l p g p represents the high-level feature information, low-level feature information, and global feature information of the input layer, respectively. h 'and p l ' represents the output corresponding to high-level features and the output corresponding to low-level features, respectively. Conv represents convolution, BatchNormalization and ReLU activation operations, "*" represents pixel-level multiplication, and "+" represents pixel-level addition.
[0030] Preferably, the process of redesigning the loss function using PolyLoss and constructing a Poly-based cross-entropy loss function includes at least the following steps:
[0031] The cross-entropy loss function is based on (1-P) tThe Taylor expansion polynomial of is as follows:
[0032]
[0033] This loss function, by adjusting the polynomial coefficients, helps to optimize the cross-entropy loss function;
[0034] Using the Poly-1 loss formula, a hyperparameter and a line of code are introduced, namely, adjusting the coefficient of the first term to adjust it, to obtain a new cross-entropy loss function;
[0035] The cross-entropy loss function is:
[0036]
[0037] Where P t σ is the optimal hyperparameter for predicting confidence levels.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] This invention uses the Swing Transformer network structure as the backbone network to extract multi-level features; the channel recalibration module recalibrates the channel feature information of the backbone network to reduce the amount of computation; the cross-fusion module incorporates global information to progressively fuse feature information while multiplying and fusing; the Poly-based cross-entropy loss function can adjust the first term coefficient of the Taylor expansion of the BCE loss function in experiments to obtain better performance; the entire two-level decoder progressive feedback network can achieve the best performance of this structure in the salient target recognition task. Attached Figure Description
[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This is the overall network structure of the salient target recognition scheme based on the Swing Transformer of this invention;
[0042] Figure 2 This is a network structure diagram of the channel recalibration module of the present invention;
[0043] Figure 3 This is a detailed structural diagram of the cross-fusion module of the present invention;
[0044] Figure 4 The experimental results of the comparative experiment of this invention are shown in the PR curve.
[0045] Figure 5 The experimental results of the comparative experiment of this invention are shown in the FM curve.
[0046] Figure 6 This is a visual comparison result of the comparative experiment of the present invention. Detailed Implementation
[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0048] Example 1:
[0049] This embodiment discloses a salient object recognition method based on Swin Transformer. Swin Transformer acts as an encoder to extract hierarchical features, an attention mechanism bridges the gaps between features at different levels, global information guides the detection of salient regions, and feedback information refines the boundaries of salient objects. Specifically, the Swin Transformer encoder first extracts multi-level features, then a channel recalibration module optimizes intra-layer channel features. A feature fusion module, guided by global information, achieves inter-layer feature fusion. To clarify blurred boundaries, the second-stage feature extraction refines edges under the guidance of feedback information.
[0050] Please see Figure 1-3 A salient target recognition method based on Swin Transformer includes at least the following steps:
[0051] The Swin Transformer is used as the backbone of the network architecture to extract multi-level feature maps from images. The core idea of the Swin Transformer is to combine the powerful modeling Transformer structure with important visual signal priors, including hierarchy, locality, and translation invariance. A key design feature of the Swin Transformer is shifted windows, which significantly reduces computational complexity, allowing computational complexity to increase linearly with the size of the input image. Furthermore, unlike traditional sliding windows, the non-overlapping window design is more hardware-friendly, resulting in faster operation. For a detailed diagram of the overall network structure, please refer to [link to network architecture diagram]. Figure 1 ;
[0052] A channel calibration module is established to reduce subsequent computation and recalibrate channel feature information, thereby optimizing both high-level and low-level feature information. During application, the channel calibration module copies the feature information obtained from the backbone network, fuses the two sets of feature information in parallel, and uses a convolution to change the number of channels, reducing the computational burden of subsequent modules and obtaining optimized channel feature information for subsequent feature fusion. The network structure of this module is as follows: Figure 2 As shown;
[0053] A cross-fusion module is established to progressively fuse feature information by incorporating global information while multiplying and fusing features. The cross-fusion module is a three-input, two-output module. It multiplies and fuses the features of two adjacent layers, and then adds the sum of the corresponding high-level or low-level feature maps and the global feature map as residual blocks to the corresponding high-level or low-level interfaces. The cross-fusion module can effectively fuse key feature information of adjacent layers, effectively distinguish foreground and background information, fuse corresponding regional feature information of neighboring layers, better preserve block-shaped feature information, and more accurately optimize the integrity of significant feature regions.
[0054] The feature information was optimized through the channel recalibration module, resulting in richer feature information. These optimized high-level features contain rich semantic information but lack boundary detail; low-level features contain rich detail but also have a lot of background noise. How to properly process and fuse this optimized feature information becomes a new challenge. Neighboring layer features have the highest similarity, while global features are rich in semantic information. Skillfully fusing neighboring layer and global feature information can better obtain the desired key features. The application of the cross-fusion module includes at least the following steps:
[0055] The high-level and low-level feature information optimized by the channel calibration module are sampled into images of the same resolution, and then the two feature maps are multiplied and fused to obtain a fused feature map.
[0056] The high-level or low-level feature maps are sampled to their respective resolutions. The high-level and low-level feature maps of the initial input are convolved, normalized, and activated, and then added as residual blocks to the fused features.
[0057] The global feature map is sampled to the resolution of both the high-level and low-level feature maps, and is also added as a residual block to the fused feature.
[0058] The high-level and low-level feature maps are respectively processed by convolutional normalization activation to obtain new feature maps of high-level and low-level features, which are then used for feature fusion in the next stage.
[0059] The structural formula of the cross-fusion module is as follows, and its specific structural diagram is as follows. Figure 3 As shown:
[0060] p h '=Conv(Conv(Conv(Conv(p h ))*Conv(Conv(p l )))+Conv(p h )+Conv(p g ))
[0061] p l '=Conv(Conv(Conv(Conv(p h ))*Conv(Conv(p l )))+Conv(p l )+Conv(p g ))
[0062] Where, p h p l p g p represents the high-level feature information, low-level feature information, and global feature information of the input layer, respectively. h 'and p l ' represents the output corresponding to high-level features and the output corresponding to low-level features, respectively. Conv represents convolution, BatchNormalization and ReLU activation operations, "*" represents pixel-level multiplication, and "+" represents pixel-level addition.
[0063] The loss function was redesigned using PolyLoss, and a Poly-based cross-entropy loss function was constructed to adjust the first term coefficient of the Taylor expansion of the BCE loss function in experiments to obtain better performance.
[0064] The process of redesigning the loss function using PolyLoss and constructing a Poly-based cross-entropy loss function includes at least the following steps:
[0065] The cross-entropy loss function is based on (1-P) t The Taylor expansion polynomial of is as follows:
[0066]
[0067] This loss function, by adjusting the polynomial coefficients, helps to optimize the cross-entropy loss function;
[0068] The low-order polynomials contribute a large portion of the gradient during training, and their coefficients are related to the prediction confidence Pt. Experiments show that the adjustment based on the coefficient of the first term has the greatest impact, while the subsequent higher-order terms have little impact on the results.
[0069] A simple and effective Poly-1 loss formula was adopted, and a hyperparameter and a line of code were introduced, namely, adjusting the coefficient of the first term to adjust it, so as to obtain a new cross-entropy loss function.
[0070] The cross-entropy loss function is:
[0071]
[0072] Where P t For predicting confidence, σ is the optimal hyperparameter;
[0073] A two-stage feedback progressive network is constructed using Swin Transformer as the backbone of the network architecture, a channel recalibration module, a cross-fusion module, and a Poly-based cross-entropy loss function to complete the identification of salient targets.
[0074] In summary, a novel network framework for salient object recognition based on the Swin Transformer is proposed. First, to address the issue of insufficient feature information extracted by the backbone network, the Swin Transformer is used instead of the traditional ResNet and VGG backbone networks. Then, a channel recalibration module is proposed to integrate the feature information extracted by the backbone network. To bridge the differences between features at different levels, a cross-fusion module is proposed to adaptively integrate features from neighboring layers and global features, suppressing background noise and supplementing missing feature information. To further obtain finer details, a Poly-based loss function is introduced, and a better network performance is achieved by adjusting the first term coefficient of the Taylor expansion of the BCE loss function during training.
[0075] The entire network framework demonstrates excellent feature extraction capabilities, exhibiting stability and efficiency across a variety of challenging scenarios. Experimental results on five popular public datasets show that the proposed network achieves state-of-the-art performance among current state-of-the-art methods.
[0076] Example 2:
[0077] This embodiment further discloses a corresponding ablation experiment based on the above embodiments;
[0078] A series of ablation experiments were designed to verify the impact of different modules on performance. The table below shows the impact of each module on the results of salient object recognition. A series of ablation experiments were conducted on three widely used datasets for salient object recognition: DUTS-TE, HKU-IS, and PASCAL-S. The evaluation metrics used were Smeasure (Sm), Mean Average Error (MAE), and max Fmeasure (maxF). The same equipment and experimental environment were used throughout the experiments. The data in the first row of Table 1 shows the baseline performance metrics, using ResNet50 as the backbone.
[0079] The first improvement was changing the backbone to a Swin Transformer and adjusting related hyperparameters through extensive experiments to achieve good performance. This improvement was crucial and yielded the greatest performance boost. The second row of Table 1 shows the effect after changing the ResNet backbone to a Swin Transformer backbone based on the baseline. Next, the features were optimized through a channel calibration module. This module replicated the features extracted by the backbone and cascaded and fused them, then reduced the number of channels through convolution to obtain optimized feature information. This module also contributed to the overall performance improvement, as shown in the third row of Table 1. An improvement based on PolyLoss was made to the overall loss function; the results of the PolyLoss improvement are shown in the fourth row of Table 1. Finally, all modules were combined to obtain the final model performance, as shown in the last row of the table. It can be seen that each module is crucial to the entire deep learning framework, and each module has a positive impact on the model's performance, demonstrating that each module in the entire deep learning framework is indispensable.
[0080] Table 1 Ablation Experiment: Effect of Modules
[0081]
[0082] Because a PolyLoss-based loss function was designed, and a simple and effective Poly-1 loss formula was adopted, it only introduces one hyperparameter and one line of code, namely adjusting the first term (1-P). tThe hyperparameter σ was adjusted using a coefficient. This hyperparameter needed to be adjusted during the experiments, and a series of ablation experiments were conducted to obtain an optimal σ. The experiments were performed using the same equipment and under the same environment. Four sets of ablation experiments were conducted with σ values of 0.2, 0.4, 0.8, and 1.2, and the results are shown in Table 2. It can be seen that the network model performed best when the value of σ was 0.4.
[0083] Table 2: Ablation Experiment: Effect of Different σ on the Experiment
[0084]
[0085] Example 3:
[0086] This embodiment further discloses a performance comparison experiment based on the above embodiments;
[0087] The proposed network performance was evaluated on five popular datasets, including ECSSD (1000 images), Pascal-S (850 images), DUT-OMRON (5168 images), HKU-IS (4447 images), and DUT (15,572 images).
[0088] All datasets were manually labeled at the pixel level for quantitative evaluation.
[0089] DUTS is currently the largest SOD dataset, consisting of 10,553 training images (DUTS-TR) and 5,019 test images (DUTS-TE). Since most previous salient object recognition works used DUTS-TR as the training dataset and all other datasets as test datasets, this case study adopts the same dataset partitioning method.
[0090] To demonstrate the effectiveness of the proposed network architecture implementation algorithm, it was compared with 10 state-of-the-art SOD algorithms, including PFSNet, DPNet, GateNet, GCPANet, MINet, PAKRNet, U2NET, LDFNet, F3Net, and PGNet.
[0091] For fair comparison, all saliency plots used were provided by the authors and evaluated using the same evaluation code.
[0092] As shown in Table 3, compared with other methods, the proposed method achieves excellent scores on six metrics across five datasets. This demonstrates the superior performance of the proposed network architecture. Furthermore, Figure 4 and Figure 5The precision-recall (PR) and FM (Fast Moving Average) curves of the above algorithm are shown on five datasets, which can evaluate the overall performance of the model. These curves demonstrate that our proposed network performs at a state-of-the-art level across different thresholds, indicating that our method has a strong ability to detect salient regions and generate accurate saliency maps.
[0093] Performance was compared with nine state-of-the-art methods on five datasets. Model performance was measured using metrics such as Sm (higher is better), MAE (lower is better), wfm (higher is better), maxF (higher is better), and maxE (higher is better). The top three methods for each dataset in the table are shown in bold for each metric. The backbone network for "Ours*" is swin-base-224, and the backbone network for "Ours" is swin-base-384. Our model ranked among the top on all five most mainstream datasets.
[0094] Table 3 compares the performance with 10 state-of-the-art methods on five datasets.
[0095]
[0096]
[0097] PR and FM curves were then plotted to compare the performance of various methods. The red curve represents our method; the solid red line represents the Swin-base-384 backbone network, and the dashed red line represents the Swin-base-224 backbone network. It can be seen that our method performs well among all state-of-the-art (SOTA) methods.
[0098] Finally, in order to evaluate the proposed network, Figure 6The figure visualizes the saliency maps generated by the self-optimizing feedback network model and other methods. It can be observed that the proposed self-optimizing feedback network not only highlights salient target regions but also effectively suppresses background noise. It excels at handling various challenging scenarios. The first column of the figure is the original image, the second column is the ground truth, the third column is the visualization result of our method, and the following columns 4-11 are a comparison of visualization results with some current state-of-the-art (SOTA) methods. Similarly, the results of these methods are provided directly by the authors or by their original training models. We use a unified evaluation code to evaluate the saliency prediction maps published by each method. It can be seen that rows 1-4 of the figure show a prediction for some occluded scenes; our method can predict the occluded parts well and exclude them from the background region. Rows 5-9 show the detailed processing; our method also has advantages in both the overall prediction and the handling of some edge details. The following rows show the prediction visualization results for cluttered scenes and scenes with background interference; it can also be seen that our method has advantages over other SOTA methods. In summary, our proposed method has better robustness and better performance in salient target recognition in complex scenes compared to current methods. More importantly, the network achieved this effect without any post-processing.
[0099] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A salient target recognition method based on Swing Transformer, characterized in that: At least the following steps are included: The Swing Transformer is used as the backbone of the network architecture to extract multi-level feature maps from images. Establish a channel calibration module to recalibrate channel feature information and reduce the computational load of subsequent feature fusion; A cross-fusion module is established to progressively fuse multi-level feature information by adding global information while multiplying and fusing. The cross-fusion module is a three-input, two-output module. The cross-fusion module multiplies and fuses the features of two adjacent layers, and then adds the corresponding high-level or low-level feature map and global feature map as residual blocks to the corresponding high-level or low-level interface. The loss function was redesigned using PolyLoss, and a Poly-based cross-entropy loss function was constructed to adjust the first term coefficient of the Taylor expansion of the BCE loss function in experiments to obtain better performance. A two-stage decoder progressive feedback network is constructed using the Swin Transformer as the backbone of the network architecture, a channel recalibration module, a cross-fusion module, and a Poly-based cross-entropy loss function to complete the recognition of salient targets. The application of the channel calibration module includes at least the following steps: Copy the feature information obtained from the backbone network; The two sets of feature information are combined in parallel and fused. By using a convolution to change the number of channels, the computational load of subsequent modules can be reduced. Optimized channel feature information is obtained for subsequent feature fusion; The application of the cross-fusion module includes at least the following steps: The high-level and low-level feature information optimized by the channel calibration module are sampled into images of the same resolution, and then the two feature maps are multiplied and fused to obtain a fused feature map. The high-level or low-level feature maps are sampled to their respective resolutions. The high-level and low-level feature maps of the initial input are convolved, normalized, and activated, and then added as residual blocks to the fused features. The global feature map is sampled to the resolution of both the high-level and low-level feature maps, and is also added as a residual block to the fused feature. The high-level and low-level feature maps are respectively convolutionally normalized and activated to obtain new high-level and low-level feature maps for the next stage of feature fusion.
2. The salient target recognition method based on Swing Transformer according to claim 1, characterized in that: The structural formula of the cross-fusion module is as follows: in, , , These represent the high-level feature information, low-level feature information, and global feature information of the input layer, respectively. and These represent the outputs corresponding to high-level features and low-level features, respectively. This represents convolution, BatchNormalization, and ReLU activation operations. "Indicates pixel-level multiplication operation," "" indicates pixel-level addition.
3. The salient target recognition method based on Swing Transformer according to claim 1, characterized in that: The process of redesigning the loss function using PolyLoss and constructing a Poly-based cross-entropy loss function includes at least the following steps: Based on the cross-entropy loss function The Taylor expansion polynomial is as follows: This loss function, by adjusting the polynomial coefficients, helps to optimize the cross-entropy loss function; Using the Poly-1 loss formula, a hyperparameter and a line of code are introduced, namely, adjusting the coefficient of the first term to adjust it, to obtain a new cross-entropy loss function; The cross-entropy loss function is: in To predict confidence levels, These are the optimal hyperparameters.