A method for detecting salient objects under different noise scenes

Abstract

This invention discloses a salient target detection method for different noise scenarios, proposing a general SOD framework tailored to different weather conditions, including a specific encoder and a replaceable decoder. To address the noise interference problem under complex weather conditions, the framework introduces a noise indicator fusion module that uses semantic features and noise indicators as dual inputs to guide the encoder in learning robust feature representations, thereby achieving adaptive processing of noise caused by various weather conditions. This invention effectively overcomes the performance limitations of traditional SOD schemes in complex weather scenarios, providing reliable technical support for improving the detection accuracy and stability of SOD tasks under different weather conditions.

Description

Technical Field

[0001] This invention belongs to the field of salient object detection in optical images based on deep learning technology, and specifically relates to a method for salient object detection in different noise scenarios. Background Technology

[0002] Saliency object detection (SOD) is a crucial research area in computer vision, focusing on automatically identifying and locating the regions in an image that most effectively capture human visual attention. It has long been a research focus, and with advancements in research, the field has achieved leaps from unimodal RGB SOD to multimodal paradigms, including RGB-Thermal, RGB-Depth, and even trimodal RGB-Depth-Thermal SOD. These advancements have enhanced the generalization ability and scene adaptability of SOD models. However, despite these significant progresses, most existing SOD methods assume that the input image is under ideal visual conditions (i.e., noise-free or low-noise environments), neglecting the impact of weather-induced noise on segmentation accuracy in real-world scenes.

[0003] Existing typical strategies for processing images with varying weather noise levels fall into two categories: one is to train and test using images with diverse weather noise levels without differentiation. This strategy completely ignores the inherent variations between different weather noise types, limiting the model's adaptability under complex conditions. The other is to train a dedicated model for each weather type. While tailored to specific noise patterns, this requires storing multiple sets of model parameters, inevitably increasing computational costs. Therefore, this invention proposes a general strategy to guide the encoder to learn noise-related discriminative feature representations through noise metrics, while ensuring compatibility with a wide range of existing decoders. A single encoder simplifies the architecture, while explicit noise-aware conditioning enhances cross-weather type discriminative capabilities. Summary of the Invention

[0004] To address the performance degradation issue in existing salient target detection tasks under complex weather conditions, this invention provides a salient target detection method for different noise scenarios. The general framework is decomposed into core components: a specific encoder and a replaceable decoder. Our design focuses on the specific encoder. Specifically, a general SOD framework tailored to different weather conditions is proposed, including a specific encoder and a replaceable decoder. The introduced Noise Indicator Fusion Module (NIFM) takes semantic features and noise indicators as dual inputs and guides the encoder to learn robust feature representations, thereby achieving adaptive processing of noise caused by various weather conditions.

[0005] A salient target detection method for different noise scenarios, the basic steps of which are as follows:

[0006] Step 1: Dataset creation. Establish the image dataset needed for network training in Step 2.

[0007] Step 2: Construct a general noise network that includes a ResNet-50 backbone network with a noise indicator fusion module embedded and replaceable decoders.

[0008] Step 3: Use the training set in the dataset as input to train the constructed general noise network.

[0009] Step 4: The test set in the dataset is processed by the trained network to obtain the saliency prediction map.

[0010] In one possible implementation, step 1 is specifically described as follows:

[0011] Step 1.1: To ensure scene diversity, acquire several existing datasets with clean images and corresponding annotations. First, retain outdoor scenes from the datasets that are compatible with severe weather noise; then, randomly divide the retained RGB images into two parts in a 7:3 ratio, which will serve as the base images for the subsequent synthetic training set and synthetic test set, respectively.

[0012] Step 1.2: Synthesize scene images

[0013] An image-based style transformation library was selected to simulate weather noise, including fog, rain, snow, darkness, overexposure, fog + rain, fog + snow, and rain + snow. For the synthesized training set, each RGB image was randomly processed into 2-5 noise images with different noise types, while the original RGB images were randomly retained. For the synthesized test set, a portion of the RGB images were randomly selected to generate noise images with random noise types, while the unselected original RGB images were retained.

[0014] Step 1.3: Acquire real-world scene images

[0015] Real-world images were collected as the real test set. Specifically, several candidate images with realistic weather noise were first obtained. Then, the candidate images were further filtered based on image content, image quality, and noise type, requiring images to retain a clear foreground subject and clear visual semantics, and ensuring coverage of different scenes under each weather type; simultaneously, overly blurry or overexposed images were removed, resulting in test images with balanced weather types and diverse scenes. Next, multiple annotators were selected to saliency-label the images. Based on the decision records of all annotators, common objects identified by at least half of the annotators were designated as final salient objects. Next, pixel binary masks were generated for the final salient objects. Finally, the pixel binary mask images with the final salient objects and the corresponding original RGB images constituted the real test set.

[0016] In one possible implementation, step 2 is specifically described as follows:

[0017] Step 2.1: Design a general SOD framework, i.e., a general noise network, for different weather conditions. This framework includes a specific encoder and a replaceable decoder. To handle different weather noises, a one-hot encoding is introduced as a noise indicator to represent different noise types, and a noise indicator fusion module is designed. The noise indicator fusion module accepts semantic features and noise indicators as dual inputs and inserts them between successive stages of the encoder, embedding weather-aware priors through adaptive feature modulation.

[0018] Step 2.2: Extract multi-scale semantic features using a specific encoder.

[0019] The encoder employs a ResNet-50 backbone network embedded in a noise indicator fusion module NIFM, given an input image with a known noise type. ∈ We construct a one-hot code, i.e., a noise indicator T, where different noise indicators correspond to different predefined noise types. Assuming there are M different noise types, then the noise indicator T ∈ It is a vector of length M, where the element corresponding to the noise type of the input image is set to 1, and the other elements are set to 0. Then the input image... Inputting a noise indicator T into the ResNet-50 backbone network to extract multi-scale features A noise indicator fusion module is inserted between every two adjacent stages of the ResNet-50 backbone network. For the first noise indicator fusion module, the input consists of two parts: semantic features extracted from the first stage of the backbone network. And a pre-constructed noise indicator T. First, semantic features. The vector is compressed into a one-dimensional vector using a global average pooling (GAP) layer, and then concatenated with a noise indicator T along the channels to obtain cascaded features. It can be expressed as the following formula:

[0020] =

[0021] In the formula, This refers to the operation of splicing by channel.

[0022] Next, cascade features Two fully connected layers (fc) are fed in to learn the noise indicator weights. The weight calculation formula is:

[0023] =

[0024] In the formula: This represents the sigmoid operation.

[0025] Finally, the learned weights Modulation applied to the backbone network, thereby incorporating anti-interference conditions and generating enhanced output characteristics in the first stage. .

[0026]

[0027] In the formula: ⊗ indicates multiplication by channel.

[0028] For the nth noise indicator fusion module (where n=2,3,4), the input is semantic features. (i.e., the features extracted from the nth stage of the backbone network) and the noise indicator T. Therefore, multi-scale semantic feature extraction yields... ,in It is by Features obtained only through stage 5.

[0029] Step 2.3: Use a decoder to map multi-scale semantic features into saliency maps.

[0030] The decoder outputs five scale features from the ResNet-50 backbone network. As input, the final output is a single-channel significance prediction map.

[0031] In one possible implementation, step 3 is specifically described as follows:

[0032] The general-purpose noisy network was trained using the training set constructed in step 1. First, the resolution of the input images in the dataset was standardized. The Adam optimizer was used to update the network parameters, and a StepLR learning rate scheduling strategy was employed to ensure stable convergence.

[0033] The loss function of a general noise network is expressed as follows:

[0034] + +

[0035] in, , and These represent BCE loss, SSIM loss, and IoU loss, respectively.

[0036] In one possible implementation, step 4 is specifically described as follows:

[0037] Input the test set (containing two types of data samples, synthetic test set and real test set) constructed in step 1 into the trained general noise network; through the inference operation of the trained general noise network, output a saliency prediction map that corresponds one-to-one with each sample in the test set.

[0038] The beneficial effects of this invention are as follows:

[0039] This invention proposes a universal salient object detection (SOD) framework tailored to different weather conditions, comprising two core components: a specifically designed encoder and a replaceable decoder. To address noise interference in complex weather conditions, the framework introduces a Noise Adaptive Feature Fusion (NIFM) module. This module uses semantic features and noise indicators as dual inputs, guiding the encoder to learn robust feature representations, thereby achieving adaptive processing of noise caused by various weather factors. In summary, this universal SOD framework, through structured design and functional adaptation to the NIFM module, effectively overcomes the performance limitations of traditional SOD schemes in complex weather scenarios, providing reliable technical support for improving the detection accuracy and stability of SOD tasks under different weather conditions. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0041] Figure 1 This is an overall architecture diagram of the general network model in an embodiment of the present invention.

[0042] Figure 2 This represents a quantitative result of 100% of the data volume in the embodiments of the present invention.

[0043] Figure 3 This represents 50% of the quantitative results from the embodiments of the present invention.

[0044] Figure 4 This represents 30% of the quantitative results from the embodiments of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0046] This application provides a method for detecting salient targets in different noise scenarios, the basic steps of which are as follows:

[0047] Step 1: Dataset creation. Establish the image dataset needed for network training in Step 2.

[0048] Step 2: Construct a general noise network that includes a ResNet-50 backbone network with a noise indicator fusion module embedded and replaceable decoders.

[0049] Step 3: Use the training set from the dataset in Step 1 as input to train the general noise network constructed in Step 2.

[0050] Step 4: The test set from the dataset in Step 1 is processed by the trained general noise network to obtain the saliency prediction map.

[0051] In one possible implementation, step 1 is specifically described as follows:

[0052] Step 1.1: To ensure scene diversity, several existing datasets with clean images and corresponding annotations were acquired, including DUTS, DUT-O, ECSSD, HKU-IS, and VT5000. First, outdoor scenes compatible with severe weather noise were retained from the datasets; then, the retained RGB images were randomly divided into two parts at a 7:3 ratio, serving as the base images for the subsequent synthetic training and test sets, respectively.

[0053] Step 1.2: Synthesize scene images

[0054] An existing image-based style transformation library was selected to simulate weather noise, including fog, rain, snow, darkness, overexposure, fog + rain, fog + snow, and rain + snow. In this embodiment, the style transformation library used is imgaug. By calling the corresponding functions in the imgaug.augmenters.weather module, random weather transformations are performed on the input image to achieve realistic image degradation. The visual effect of each type of weather noise can be adjusted by controlling the intensity parameter. For example, to simulate rain noise, the original image is first read, then a rain style enhancement is defined, and finally the enhancement effect is applied and the image is saved. For the synthesized training set, each RGB image is randomly processed into 2-5 noise images with different noise types, while the original RGB images are randomly retained. For the synthesized test set, a portion of the RGB images are randomly selected to generate a single noise image with a random noise type, while the unselected original RGB images are retained.

[0055] Step 1.3: Acquire real-world scene images

[0056] Domain shifts also exist between clean and noisy images, as well as between synthetic and real noise. To improve data generalization, 554 real-world images were collected as the real test set. Specifically, over 1,500 candidate images with real weather noise were first searched from several common search engines such as Google, Baidu, and Bing. Then, the candidate images were further filtered based on image content, image quality, and noise type, requiring images to retain clear foreground subjects and have explicit visual semantics, and ensuring coverage of different scenes such as cities, rural areas, highways, and natural landscapes under various weather types; simultaneously, excessively blurry or overexposed images were removed, resulting in over 600 test images with balanced weather types and diverse scenes. Seven annotators were then selected to saliency-label the images. Based on the decision records of all annotators, common objects identified by at least half of the annotators were designated as final salient objects. Next, pixel binary masks were carefully generated for the final salient objects using Photoshop. Finally, the 554 pixel binary mask images with the final salient objects, along with their corresponding original RGB images, constituted the real test set.

[0057] In one possible implementation, step 2 is specifically described as follows:

[0058] Step 2.1: Design a general SOD framework, i.e., a general noise network, for different weather conditions. This framework includes a specific encoder and a replaceable decoder. To handle different weather noises, a one-hot encoding is introduced as a noise indicator to represent different noise types, and a noise indicator fusion module is designed. The noise indicator fusion module accepts semantic features and noise indicators as dual inputs and is interpolated between successive stages of the encoder, embedding weather-aware priors through adaptive feature modulation. Crucially, the proposed specific encoder maintains compatibility with mainstream SOD decoders.

[0059] Step 2.2: Extract multi-scale semantic features using a specific encoder.

[0060] The encoder employs a ResNet-50 backbone network embedded in a noise indicator fusion module NIFM, given an input image with a known noise type. ∈ Construct a one-hot code, i.e., a noise indicator T. It uniquely encodes the noise type, with different noise indicators corresponding to different predefined noise types. Assuming there are M different noise types, then the noise indicator T∈ It is a vector of length M, where the element corresponding to the noise type of the input image is set to 1, and the other elements are set to 0. For example, the noise indicator for a rain image is [0,1,0,0,0,0,0,0,0], while the noise indicator for a snow image is [0,0,1,0,0,0,0,0,0]. Then the input image... Inputting a noise indicator T into the ResNet-50 backbone network to extract multi-scale features .like Figure 1 As shown in (a), a noise indicator fusion module is inserted between every two adjacent stages of the ResNet-50 backbone network. For the first noise indicator fusion module, the input consists of two parts: semantic features extracted from the first stage of the backbone network. And a pre-built noise indicator T. For example... Figure 1 As shown in (b), firstly, semantic features The vector is compressed into a one-dimensional vector using a global average pooling (GAP) layer, and then concatenated with a noise indicator T along the channels to obtain cascaded features. It can be expressed as the following formula:

[0061] =

[0062] In the formula, This refers to the operation of splicing by channel.

[0063] Next, cascade features Two fully connected layers (fc) are fed in to learn the noise indicator weights. This weight vector encodes the degree to which the backbone features need adjustment to counteract a specific type of noise, represented by T. The weight calculation formula is as follows:

[0064] =

[0065] In the formula: This represents the sigmoid operation.

[0066] Finally, the learned weights Modulation applied to the backbone network, thereby incorporating anti-interference conditions and generating enhanced output characteristics in the first stage. .

[0067]

[0068] In the formula: ⊗ indicates multiplication by channel.

[0069] It can be seen that, Features implemented The noise type is switched. The same operational logic extends to all subsequent noise indicator fusion modules. For the nth noise indicator fusion module (where n=2,3,4), the input is semantic features. (i.e., the features extracted in the nth stage of the backbone network) and the noise indicator T. Thus, for the remaining stages, the noise type features and their associated indicator weights are derived progressively. Therefore, multi-scale semantic feature extraction yields... ,in It is by Features obtained only through stage 5.

[0070] Step 2.3: Use a decoder to map multi-scale semantic features into saliency maps.

[0071] Mainstream backbone-based SOD methods can generally be decomposed into two core parts: an encoder and a decoder. The encoder is designed to extract multi-scale semantic features, while the decoder is distinguished by different architectural designs (e.g., pyramid feature fusion, cross-scale attention, or residual refinement modules), and its task is to progressively map these hierarchical semantic features to dense pixel-level saliency maps. This invention focuses on designing a general framework rather than a specific network structure. The decoder uses five scale features output from the ResNet-50 backbone network. As input, the final output is a single-channel significance prediction map.

[0072] In one possible implementation, step 3 is specifically described as follows:

[0073] The general noise network is trained using the training set constructed in step 1. First, the resolution of all input images in the dataset is adjusted to 384×384 (for standardization). The Adam optimizer is then used... =0.9 and The network parameters are updated using γ = 0.999, with an initial learning rate of 0.001. To ensure stable convergence, a StepLR learning rate scheduling strategy is adopted, with the learning rate decreasing by γ = 0.2 every 40 epochs, and training for 52 epochs.

[0074] The loss function of a general noise network is expressed as follows:

[0075] + +

[0076] in, , and These represent the BCE loss, SSIM loss, and IoU loss, respectively. The calculation process for each loss function is as follows:

[0077]

[0078]

[0079]

[0080] in and The significance plot and the truth plot represent the predictions. , and , They are and The mean and standard deviation, It is their covariance. =0.012 and =0.032 is used to avoid the denominator being 0.

[0081] In one possible implementation, step 4 is specifically described as follows:

[0082] Input the test set (containing two types of data samples, synthetic test set and real test set) constructed in step 1 into the trained general noise network; through the inference operation of the trained general noise network, output a saliency prediction map that corresponds one-to-one with each sample in the test set.

[0083] To verify the functional value of the NIFM module, quantitative evaluation metrics were further calculated for the generated saliency prediction map. These metrics mainly fall into two categories: core evaluation metrics, including the error evaluation metric MAE (measuring the mean absolute error between the predicted saliency map and the true saliency annotation) and the similarity evaluation metric SM (measuring the structural similarity between the two); and two key consistency evaluation metrics—the F-measure (FM) and the E-measure (EM). The F-measure (FM) measures the overlap and consistency between the saliency prediction results and the true annotations by using the harmonic average of precision and recall. It includes three different calculation forms: FM_adp (adaptive threshold F-measure), FM_mean (multi-threshold average F-measure), and FM_max (maximum threshold F-measure). The E-measure (EM) includes EM_adp, EM_mean, and EM_max, which correspond to the same threshold calculation logic as FM_adp, FM_mean, and FM_max, respectively, and are used to further measure the consistency between the prediction results and the true annotations from an error perception perspective. Another type is the difference index starting with "Δ" (ΔMAE, ΔSM, ΔFM_adp, etc.), which respectively represent the difference between the corresponding core indexes in this invention and the basic model, and are used to intuitively reflect the improvement effect of this invention on model performance. Based on the calculation results and analysis of quantitative evaluation indicators, it can be clearly demonstrated that the NIFM module has a significant effect on improving the accuracy of network model saliency prediction and optimizing the detail representation of prediction graphs, effectively supporting the technical advantages of this invention.

[0084] Experimental results:

[0085] To fully validate the effectiveness of the proposed general framework, especially its integration with a specific encoder in NIFM, quantitative experiments were conducted from two perspectives: 1) the effectiveness of the proposed framework when paired with different decoder configurations; and 2) the effectiveness of the proposed framework at different dataset scales.

[0086] 1) Figure 2 In this paper, we present the results of an ensemble model integrating seven different decoders (including F3Net, RNet, MINET, AESINet, BAFSNet, DcNet, and EDN, with detailed structural configurations referenced from the original literature) on synthetic and real test sets. Among them, " "This indicates a model that uses a traditional ResNet-50 encoder, while " "" indicates models equipped with a specific encoder (integrated with the Noise Indicator Fusion Module NIFM). Quantitative analysis shows that after replacing the traditional encoder with our specific encoder, the average MAE of all seven models was reduced by 7.27% on the synthetic test set and 17.70% on the real test set. Consistent performance enhancements were also observed in other evaluation metrics besides MAE, with greater improvements on the real test set compared to the synthetic test set. This difference highlights the specific encoder's superior ability to mitigate real-world weather noise, which exhibits more complex and unstructured characteristics compared to synthetic noise.

[0087] 2) To further evaluate the stability of the proposed general framework across various dataset scales, all seven models were retrained using two reduced training subsets of the aforementioned datasets (i.e., 50% and 30% training sets). Results are shown below. Figure 3 , Figure 4 First, an intuitive trend is observed: reducing the amount of training data leads to a corresponding decrease in the average performance across all seven datasets. For example, when the training set is reduced from 100% (original) to 50% and 30%, the average MAE on the real test set decreases from 0.0198 to 0.0230 and 0.0260, respectively. Notably, even with the performance degradation due to reduced training data, the integration of our Noise Index Fusion Module (NIFM) consistently improves the segmentation accuracy of both synthetic and real-world test subsets, regardless of the training set size (50% or 30%). This sustained improvement highlights NIFM's ability to effectively utilize available training data to learn noisy robust cues, thereby mitigating the adverse effects of data scarcity. These findings confirm the feasibility and practicality of our proposed general framework, particularly in data-constrained real-world applications.

[0088] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0089] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0090] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A method for salient target detection under different noise scenarios, characterized in that, The basic steps are as follows: Step 1: Dataset creation, establishing the image dataset needed for network training in Step 2; Step 2: Construct a general noise network including a ResNet-50 backbone network with a noise indicator fusion module embedded and replaceable decoders; Step 3: Use the training set in the dataset as input to train the constructed general noise network; Step 4: The test set in the dataset is processed by the trained network to obtain the saliency prediction map.

2. The salient target detection method for different noise scenarios according to claim 1, characterized in that, The specific method for step 1 is as follows: Step 1.1: To ensure scene diversity, acquire several existing datasets with clean images and corresponding annotations; first, retain outdoor scenes in the datasets that are compatible with severe weather noise; then, randomly divide the retained RGB images into two parts in a 7:3 ratio, which will serve as the base images for the subsequent synthetic training set and synthetic test set, respectively. Step 1.2: Synthesize scene images An image-based style transformation library was selected to simulate weather noise, which included fog, rain, snow, darkness, overexposure, fog + rain, fog + snow, and rain + snow. For the synthesized training set, each RGB image was randomly processed into 2-5 noise images with different noise types, while the original RGB images were randomly retained. For the synthesized test set, some RGB images were randomly selected to generate noise images with random noise types, while the original RGB images that were not selected were retained. Step 1.3: Acquire real-world scene images Real-world images were collected as a real test set. Specifically, several candidate images with realistic weather noise were first obtained. Then, the candidate images were further filtered based on image content, image quality, and noise type, requiring images to retain clear foreground subjects and have clear visual semantics, and ensuring that different scenes were covered under each weather type. At the same time, overly blurry and overexposed images were removed, resulting in test images with balanced weather types and diverse scenes. Then, multiple annotators were selected to perform saliency annotation on the images. Based on the decision records of all annotators, common objects identified by at least half of the annotators were designated as the final salient objects. Next, pixel binary masks were generated for the final salient objects. Finally, the pixel binary mask image with the final salient object and the corresponding original RGB image constitute the real test set.

3. The salient target detection method for different noise scenarios according to claim 1, characterized in that, Step 2 is explained in the following steps: Step 2.1: Design a general SOD framework, i.e., a general noise network, for different weather conditions. The framework includes a specific encoder and a replaceable decoder. In order to handle different weather noises, a one-hot encoding is introduced as a noise indicator to represent different noise types, and a noise indicator fusion module is designed. The noise indicator fusion module accepts semantic features and noise indicators as dual inputs and inserts them between successive levels of the encoder, embedding weather-aware priors through adaptive feature modulation. Step 2.2: Extract multi-scale semantic features using a specific encoder; The encoder employs a ResNet-50 backbone network embedded in a noise indicator fusion module NIFM, given an input image with a known noise type. ∈ Construct a one-hot code, i.e., a noise indicator T. Different noise indicators correspond to different predefined noise types; assuming there are M different noise types, then the noise indicator T∈ It is a vector of length M, where the element corresponding to the noise type of the input image is set to 1, and the other elements are set to 0; then the input image... Inputting a noise indicator T into the ResNet-50 backbone network to extract multi-scale features A noise indicator fusion module is inserted between every two adjacent stages of the ResNet-50 backbone network. For the first noise indicator fusion module, the input consists of two parts: semantic features extracted from the first stage of the backbone network. And a pre-constructed noise indicator T; first, semantic features The vector is compressed into a one-dimensional vector using a global average pooling layer (GAP), and then concatenated with a noise indicator T along the channels to obtain cascaded features. It can be expressed as the following formula: = In the formula, This refers to the operation of splicing by channel; Next, cascade features Two fully connected layers (fc) are fed in to learn the noise indicator weights. The weight calculation formula is: = In the formula: Indicates the sigmoid operation; Finally, the learned weights Modulation applied to the backbone network, thereby incorporating anti-interference conditions and generating enhanced output characteristics in the first stage. ; In the formula: ⊗ indicates multiplication by channel; For the nth noise indicator fusion module, where n=2,3,4, the input is semantic features. That is, the features and noise indicator T extracted from the nth stage of the backbone network; therefore, multi-scale semantic feature extraction is used to obtain... ,in It is by Features obtained only through stage 5; Step 2.3: Use a decoder to map multi-scale semantic features into saliency maps; The decoder outputs five scale features from the ResNet-50 backbone network. As input, the final output is a single-channel significance prediction map.

4. A salient target detection method for different noise scenarios according to claim 1 or 3, characterized in that, The specific method for step 3 is as follows: The general noise network is trained using the training set constructed in step 1. First, the resolution of the input images in the dataset is unified. The Adam optimizer is used to update the network parameters. In order to ensure stable convergence, the StepLR learning rate scheduling strategy is adopted. The loss function of a general noise network is expressed as follows: + + in, , and These represent BCE loss, SSIM loss, and IoU loss, respectively.

5. A salient target detection method for different noise scenarios according to claim 4, characterized in that, Step 4 is explained in detail below: Input the test set constructed in step 1 into the trained general noise network; through the inference operation of the trained general noise network, output a saliency prediction map that corresponds one-to-one with each sample in the test set.

Images