A field small target lightweight detection method based on endogenous saliency guidance

Abstract

This invention discloses a lightweight detection method for small targets in the field based on endogenous saliency guidance, belonging to the field of computer vision and target detection technology. The method is based on a lightweight detection network. First, multi-scale feature extraction is performed on the input field scene image. Then, a foreground saliency response generation module is constructed using high-level semantic features to adaptively generate a saliency response map. Next, combined with small target sensitive features, the target region is enhanced by a saliency-guided feature enhancement module. The enhancement path or a weak enhancement / residual preservation path is adaptively selected based on the saliency response intensity, while local detail branches are introduced to compensate for edge, contour, and fine-grained structural information. Finally, the enhanced features and deep semantic features are fused and input into the detection head to achieve small target category recognition and bounding box localization. This invention does not require an external saliency model or additional saliency annotation, effectively suppresses interference from complex backgrounds, and improves the accuracy, stability, and robustness of small target detection in field environments, while also meeting the requirements of lightweight design and edge deployment.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and computer vision, and in particular to a lightweight detection method for small targets in the field based on saliency guidance, belonging to the field of target detection technology. Background Technology

[0002] In the field of computer vision, object detection is the cornerstone of many tasks, such as ecological monitoring, drone inspection, agricultural pest and disease identification, and intelligent field recognition systems. The main process of object detection involves inputting an image, extracting features through a deep neural network, and then classifying and locating the bounding boxes of objects in the image. However, detecting small objects in complex field environments presents several challenges, including the small size of the objects, which easily leads to feature attenuation during multiple convolutional downsampling processes; complex background textures and camouflage phenomena, making it difficult to distinguish the object from the background; severe interference from changes in lighting, occlusion, rain, and fog; and the limited expressive power of lightweight models when deployed on edge devices.

[0003] Existing detection methods typically employ a fixed feature interaction approach, where mid- or high-level features extracted from the backbone network directly participate in subsequent prediction calculations. However, in small target scenarios, if the feature representations do not yet possess sufficient discriminative power, having all region features participate in the interaction can introduce noise interference, thereby affecting the detection results.

[0004] Some studies have attempted to enhance the response of the target region by introducing external saliency maps or additional saliency branches, but there is a semantic inconsistency between external saliency information and the detection task, which can easily disrupt the original feature distribution and lead to a decline in detection performance.

[0005] Therefore, there is an urgent need for a new technical solution in the existing technology to achieve intrinsic saliency guidance for the detection task without introducing additional complex saliency networks, thereby enhancing the response capability of small target regions and improving detection stability and robustness. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] To address the problems of insufficient feature response for small targets, severe background interference, and lack of adaptive selection of feature enhancement methods in complex field environments, existing lightweight detection models suffer from these issues. This invention provides a lightweight field target detection method based on endogenous saliency guidance. This method generates a saliency response map through adaptive modeling of internal network features and adaptively controls the saliency enhancement path or the weak enhancement / residual preservation path based on the saliency response strength. Without introducing additional saliency annotation information or external saliency models, it achieves enhancement of the target region and suppression of the background region, thereby solving the problems of small targets being easily submerged by background noise, insufficient response, and lack of flexible adjustment of enhancement methods in complex field environments.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, this invention proposes a small target detection method based on endogenous saliency guidance. This method is based on a lightweight target detection network and utilizes a Foreground Saliency Response Generator (FSRG) and a Saliency-driven Enhancement Module (SDEM) to perform saliency modeling and feature enhancement on small target regions in complex outdoor environments. This enables the network to more accurately focus on target regions and suppress background interference under complex background conditions, thereby improving the accuracy and robustness of small target detection.

[0010] The proposed method comprises a feature extraction network, a foreground saliency response generation module, a saliency-guided feature enhancement module, and a detection prediction module. Each module executes sequentially according to the data processing flow, enabling saliency information to be passed layer by layer during network feature representation and ultimately impacting the target detection result. Unlike existing methods, the saliency information in this invention is adaptively generated from internal network features, eliminating the need for additional saliency annotation or external saliency models, thus ensuring semantic consistency between saliency modeling and the detection task. Specifically, firstly, the backbone network extracts multi-scale features from the input image to obtain feature maps at different semantic levels; then, mid-level features are used to construct a saliency response map to characterize the response intensity of potential target regions in the image; next, feature enhancement weights are adaptively adjusted based on the saliency response intensity to perform differentiated feature enhancement; finally, the enhanced features are input into the detection prediction module for classification and bounding box regression, thereby completing the task of detecting small targets in the field environment.

[0011] In this invention, a small target preferably refers to a target whose area accounts for no more than 3% of the total area of ​​the input image, or whose effective pixel area is no more than [a certain value]. The goal is to detect small targets against complex backgrounds. In one embodiment, the invalid background area in the image can account for more than 95%, thus forming a typical scenario for detecting small targets against complex backgrounds. This definition of small targets is used to illustrate the main scope of detection objects to which this invention is applicable, and can also be adjusted proportionally according to specific application scenarios.

[0012] Step 1: Feed the input image into the feature extraction network for multi-scale feature extraction. The input image is denoted as:

[0013]

[0014] Where 𝐻 represents the image height, 𝑊 represents the image width, and 𝐶 represents the number of image channels. The input image first undergoes preliminary feature extraction through convolutional layers, and then the receptive field is expanded through layer-by-layer downsampling operations. The backbone network consists of multiple convolutional modules, residual modules, and downsampling modules, obtaining feature representations at different semantic levels through layer-by-layer convolutional operations. Let the feature map of the l-th layer be:

[0015] l=1,2,⋯,L

[0016] Among them, 𝐹 l Let represent the output feature map of the l-th layer, 𝜙 l This represents the convolution transformation function. After multiple layers of convolution and downsampling operations, multi-scale feature maps can be obtained.

[0017] {𝐹1,𝐹2,𝐹3}

[0018] In this model, 𝐹1 is a shallow feature map with high spatial resolution, mainly containing rich texture information and edge details; 𝐹2 is a mid-level feature map, containing both spatial and semantic information; and 𝐹3 is a deep feature map with strong semantic expressive power, but its spatial resolution is relatively low. In the task of small target detection in outdoor environments, the input to the foreground saliency response generation module is taken from the deep semantic feature map of the interferometer network. This allows for more efficient use of high-level semantic information to establish response differences between foreground and background; small target sensitive features are extracted by a higher-resolution branch and used for subsequent saliency-guided enhancement. Here, "small target sensitive features" refers to features that are sensitive to smaller targets relative to the deep feature map. In contrast, feature branches have higher spatial resolution and can retain more target edge and local structural information.

[0019] Step 2: Construct a Foreground Satisfactory Response Generation (FSRG) module using deep feature maps. After obtaining the deep semantic feature maps, a saliency response map is generated through convolutional transformation to describe the foreground response intensity at each spatial location in the image. Specifically, the feature maps... Input the foreground saliency response generation module, and obtain the saliency response map through convolution operation:

[0020]

[0021] Among them, 𝐶𝑜𝑛𝑣 3×3 This represents a convolution operation with a kernel size of 3×3, used to extract local spatial context information; 𝐶𝑜𝑛𝑣 1x1`` represents a 1×1 convolution operation, used for channel compression and feature mapping; `` represents the Sigmoid activation function; `` represents the saliency response map. Each pixel value in the saliency response map represents the probability that the location belongs to the target region, and its value range is:

[0022]

[0023] Where 𝐻′ and 𝑊′ represent the height and width of the saliency response map, respectively. A larger saliency response value indicates that the region is more likely to belong to the target region. Each pixel value... This indicates the response intensity at the corresponding location belonging to the potential target region. To characterize the overall significance level, a significance statistic can be further defined:

[0024]

[0025] And based on this, an adaptive enhancement coefficient is constructed:

[0026]

[0027] in, To adjust the parameters. Thus, the saliency response can participate in the subsequent feature enhancement process with continuous weights without explicit supervision, thereby avoiding semantic inconsistency issues.

[0028] Step 3: Adaptively modulate the features using the Saliency-Guided Feature Enhancement Module (SDEM). Since the saliency response map already reflects the importance of potential target regions in the image, saliency information can be used to weight and modulate the feature map, making the network focus more on the target region. Specifically, the saliency response map is first scaled to match its spatial size with that of small target sensitive features, denoted as:

[0029]

[0030] Then, the saliency response map is compared with the small target sensitivity features. Element-by-element fusion is performed, and the calculation method is as follows:

[0031]

[0032] Where ⊙ represents element-wise multiplication. To make the enhancement process adaptive, let the enhancement threshold be... When the significance of the response intensity satisfies When the saliency-guided enhancement path is entered, its output is:

[0033]

[0034] When the significance of the response intensity does not meet the enhancement condition, i.e. When entering the weak enhancement or residual preservation path, its output is:

[0035]

[0036] in, It is a weak enhancement coefficient, and Through the aforementioned dual-path mechanism, when the saliency response is high, the corresponding regional features are enhanced; when the saliency response is low, only the features are weakly modulated and the original feature body is preserved, thereby avoiding the noise amplification problem caused by over-enhancement in low saliency scenarios.

[0037] Because small targets in the field typically have a small pixel ratio, weak local texture, incomplete edge contours, and are easily affected by vegetation, shadows, branches, and complex background textures, relying solely on saliency modulation results can easily weaken the edge and local structural information of small targets while enhancing the response of the foreground region. Therefore, this invention introduces a local detail branch in addition to the saliency enhancement path to perform local convolution modeling on the input features. The calculation method is as follows:

[0038]

[0039] in, This represents a depthwise convolution operation. The local detail branch complements the saliency enhancement path, where the saliency enhancement path improves the target region response and suppresses background noise, while the local detail branch compensates for target edges, contours, and fine-grained texture information that may be weakened during saliency modulation. Finally, the two types of features are fused to obtain the enhanced features:

[0040]

[0041] Through the above process, we can achieve saliency enhancement of the foreground region, noise suppression of the background region, and effective preservation of the edge structure of small targets.

[0042] Step 4: Fuse the enhanced features with deep semantic features. After completing the saliency-guided feature enhancement, in order to integrate higher-level semantic information, the enhanced features are... with deep features The process of fusion is represented as follows:

[0043]

[0044] Where 𝐶𝑜𝑛𝑐𝑎𝑡 represents the feature concatenation operation. Feature concatenation allows the network to simultaneously acquire saliency enhancement information and deep semantic information. Subsequently, a 1×1 convolution is used to perform channel compression and feature reshaping on the fused features, thus obtaining the final fused features:

[0045]

[0046] After the above processing, the fusion feature 𝐹 is obtained. final This feature includes saliency-enhanced spatial information and deep semantic information, providing richer feature representation for subsequent object detection. In the embodiments of the present invention, the foreground saliency response generation module is set on the high-level semantic feature output path in the backbone network, and the saliency-guided feature enhancement module is set on the small target sensitive feature path. After the enhanced feature is output, it enters the multi-scale fusion module together with the deep semantic features.

[0047] Step 5: Input the fused features into the detection and prediction module for target detection. The detection and prediction module includes a classification branch and a regression branch. The classification branch predicts the target class probability, and the regression branch predicts the location parameters of the target bounding box. The classification branch outputs the class probability:

[0048]

[0049] The regression branch outputs the target bounding box:

[0050] 𝐵=(𝑥,𝑦,𝑤,ℎ)

[0051] Where (x, y) represent the target center coordinates, and (x, y) represent the target bounding box width and height. Based on the above prediction results, accurate detection of small targets in the field environment can be achieved. To achieve joint optimization of the model, this invention constructs a comprehensive loss function:

[0052] 𝐿=𝐿 𝑐l𝑠 +𝐿 𝑟𝑒𝑔

[0053] Among them, 𝐿 𝑐l𝑠 Let f represent the classification loss function, 𝐿 𝑟𝑒𝑔 This represents the bounding box regression loss function. Since the saliency response is adaptively generated by the network's internal features, its optimization process is indirectly driven by the detection loss. Therefore, there is no need to introduce an additional saliency supervision term, thereby reducing training complexity and improving model stability.

[0054] Through the above steps, this invention achieves a complete process from image input, feature extraction, saliency modeling, feature enhancement to target detection and prediction. Salient information propagates hierarchically through the network and participates in feature representation, enabling the model to more accurately locate small target regions in complex background environments, thereby effectively improving the detection accuracy and robustness in small target detection tasks in outdoor environments.

[0055] (III) Beneficial Effects

[0056] Compared with existing technologies, this invention provides a lightweight detection method for small targets in the field based on detection supervision saliency guidance, which has the following beneficial effects:

[0057] 1. The foreground saliency response in this invention is adaptively generated by the mid-layer features inside the detection network, without the need for an external saliency network or additional saliency annotation information, thus ensuring semantic consistency between saliency modeling and the target detection task.

[0058] 2. This invention employs a dual-path adaptive enhancement mechanism based on the saliency of the response intensity. When the saliency is strong, the response of the target region is enhanced, while when the saliency is weak, weak enhancement or residual preservation is performed, thereby balancing enhancement capability and feature stability.

[0059] 3. The present invention sets up local detail branches to compensate for the edge, contour and fine-grained structural information that may be weakened during the saliency modulation process, which is beneficial to improving the complete representation of small targets in complex field backgrounds.

[0060] 4. This invention is compatible with lightweight target detection backbone structures, and achieves suppression of background noise regions and enhanced response to small target regions without significantly increasing the number of model parameters, making it suitable for deployment on edge devices.

[0061] 5. This invention can improve the accuracy, stability and robustness of small target detection in complex field environments, and has good prospects for engineering applications. Attached Figure Description

[0062] Figure 1 This is a flowchart of a lightweight detection method for small targets in the field based on saliency guidance according to the present invention;

[0063] Figure 2 This is a schematic diagram of the Foreground Significant Response Generation Module (FSRG) of the present invention;

[0064] Figure 3This is a schematic diagram of the saliency-guided feature enhancement module (SDEM) in this invention.

[0065] Figure 4 This is a general block diagram of a lightweight detection method for small targets in the field based on saliency guidance according to the present invention; Figure 5 This is a schematic diagram summarizing the main processing links of a lightweight detection method for small targets in the field based on endogenous saliency guided by the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0067] This invention proposes a lightweight detection method for small targets in the field based on saliency guidance. By introducing a saliency response generation module and a foreground feature enhancement module into the lightweight target detection network, the network can enhance the features of small foreground targets and suppress background noise in complex field environments, thereby improving the accuracy and robustness of small target detection.

[0068] like Figure 1 The flowchart shown illustrates a lightweight detection method for small targets in the field based on saliency guidance according to the present invention. The specific implementation process of this method includes the following steps:

[0069] A. Input field scene images are fed into a lightweight detection network for feature extraction to obtain multi-scale features. After feature extraction, mid-to-high-level semantic feature branches and small target sensitive feature branches are formed respectively for subsequent saliency response generation and feature enhancement.

[0070] Specifically, in this embodiment, the input image is first normalized and resized, and then input into a lightweight object detection backbone network for convolutional feature extraction. The backbone network preferably employs a lightweight detection backbone structure to balance detection accuracy and computational efficiency. Let the input image be... Then, after feature extraction through the backbone network, a multi-scale feature set is obtained. ,in, These are shallow, high-resolution features, primarily containing edge and texture details; It is a mid-level feature, possessing both spatial structural information and semantic information; As deep semantic features, they possess stronger category representation capabilities. In this embodiment, the mid-to-high-level semantic feature extraction branch utilizes deep semantic features. As input for foreground saliency response modeling, the small target sensitive feature extraction branch retains high-resolution local structural features for subsequent saliency-guided enhancement. Its feature extraction process can be represented as follows:

[0071]

[0072] in, Indicates the first Convolutional transformations and nonlinear mapping operations of layers.

[0073] B. Input the above deep semantic features into the foreground saliency response generation module FSRG to generate a saliency response map, and calculate the overall saliency response intensity by combining the saliency response map. Figure 2 The diagram shows the structure of the foreground salience response generation module FSRG in this invention;

[0074] Specifically, the FSRG module is used to adaptively generate a foreground saliency response map based on mid-to-high-level features. In this embodiment, the FSRG module first uses convolution operations to model the local spatial context of the input features, and then obtains the saliency response map through nonlinear activation and channel mapping. The process can be represented as: processing the input features... Convolution extracts local information, which is then processed by an activation function and then... The convolutional channels are compressed, and finally, a saliency response map is output using the sigmoid function. Its expression is:

[0075]

[0076] In the generated response map, the response value at each location represents the response intensity of that location belonging to the potential target region. To make the subsequent enhancement process adaptive, this embodiment further performs statistical analysis on the saliency response map to calculate the saliency response intensity, which is used to characterize the overall saliency level of the target region in the entire image. The saliency response intensity can be expressed as:

[0077]

[0078] in, and These represent the height and width of the saliency response map, respectively. The saliency response map is adaptively generated from internal network features, without relying on external saliency labels or a single saliency model, thus ensuring the consistency between saliency information and detection semantics.

[0079] C. Input the saliency response map and the small target sensitive features into the saliency-guided feature enhancement module SDEM. Adaptively select saliency-guided enhancement or weak enhancement / residual preservation strategy based on the saliency response intensity, and output the enhanced features. . Figure 3 The diagram shows the saliency guidance feature enhancement module SDEM structure in this invention.

[0080] Specifically, in this embodiment, the SDEM module includes a saliency modulation branch, a local detail enhancement branch, and a residual preservation branch. First, the saliency response map obtained in step B is sized to match the size of the small target sensitive feature space, denoted as . Subsequently, element-wise multiplication is used to apply the saliency response map to the sensitive features of small targets, in order to highlight salient regions and suppress background regions. Let the basic enhancement feature be... Then the foreground candidate features can be expressed as:

[0081]

[0082] in, This indicates element-wise multiplication.

[0083] like Figure 1 The condition for determining whether the node's "significant response intensity meets the enhancement condition" is yes, that is, when... When the saliency-guided enhancement path is reached, the feature is strongly enhanced through the saliency response map, and its output can be expressed as:

[0084]

[0085] in, This is the enhancement coefficient.

[0086] If the enhancement conditions are not met, that is, when When this happens, the system enters a weak enhancement or residual preservation path, retaining the main structure of the original feature distribution and performing only weak modulation. This avoids noise amplification caused by over-enhancement in low-saliency scenarios. The output at this time can be expressed as:

[0087]

[0088] in, It is a weak enhancement coefficient, and .

[0089] Meanwhile, this embodiment sets up a local detail branch in the SDEM module to perform local convolutional modeling on the original sensitive features, thereby supplementing edge, contour, and fine-grained structural information. This local detail branch can be represented as:

[0090]

[0091] in, This represents a depthwise convolution operation. Finally, the saliency modulation result, local detail result, and residual information are fused to obtain the enhanced features. It can be represented as:

[0092]

[0093] With the above structure, the SDEM module can enhance the response of the foreground target when the saliency is high and maintain the stability of network features when the saliency is low, thus balancing enhancement capability and robustness.

[0094] D. Enhance features The input is then processed in the subsequent multi-scale feature fusion and detection prediction module, where it works together with deep semantic features to complete target detection and output the category and bounding box location of small targets in the field. Figure 4 The diagram shows the overall block diagram of a lightweight detection method for small targets in the field based on saliency guidance according to the present invention;

[0095] Specifically, in this embodiment, Figure 4 The overall block diagram illustrates the embedding positions of the FSRG and SDEM modules in the lightweight detection network. The FSRG module is positioned after the high-level semantic feature output of the backbone network to generate a unified foreground saliency response map; the SDEM module is positioned on the small target sensitive feature path to perform saliency-guided enhancement of high-resolution features. (Enhanced Features) After output, the data enters a multi-scale feature fusion module, where it is concatenated and integrated with deep semantic features to obtain fused features that combine spatial detail information and deep semantic information. This process can be represented as:

[0096]

[0097] Subsequently, the fused features are further subjected to channel compression and feature integration to obtain the final fused features:

[0098]

[0099] in, This indicates a feature concatenation operation. Subsequently, the fused features are input into the detection head for prediction. The detection head includes a classification branch and a regression branch. The classification branch outputs the target class probability, and the regression branch outputs the target bounding box parameters, ultimately yielding the detection result. Therefore, this embodiment achieves efficient detection of small targets in complex outdoor environments through a method of "generating saliency responses from high-level semantics—performing saliency enhancement on sensitive features of small targets—outputting detection results after multi-scale fusion."

[0100] E. During model training, a detection task-driven approach is used to optimize the entire network end-to-end, enabling the generation of saliency responses and the target detection process to converge in tandem.

[0101] Specifically, in this embodiment, during training, input samples are fed into the aforementioned network structure. The network outputs target class probabilities and bounding box parameters, and classification and regression losses are constructed based on the detection results to jointly optimize the entire network parameters. Since the saliency response map is adaptively generated from the internal features of the detection network, there is no need to introduce an external saliency supervision branch, thereby reducing training complexity. Through this end-to-end training method, the FSRG module can gradually learn the differences between the target region and the background region, and the SDEM module can learn feature enhancement strategies at different saliency levels, thus enabling the entire detection network to have a stronger ability to discriminate small targets in complex backgrounds.

[0102] In this embodiment, field scene image data was used to verify the invention. The basic detection network adopted a lightweight target detection structure, and the input images were trained and tested after being standardized in size. Experimental results show that, while maintaining the lightweight characteristics, the introduction of a foreground saliency response generation module and a saliency-guided feature enhancement module enhances the model's response capability to small target regions in complex backgrounds, and the detection accuracy is improved compared to the baseline model, indicating that the solution described in this invention is feasible and effective.

[0103] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lightweight target detection method for small targets in the field based on endogenous saliency guidance, the method being based on the following modules: a lightweight target detection network, a foreground saliency response generation module (FSRG), a saliency-guided feature enhancement module (SDEM), a multi-scale feature fusion module, and a detection prediction module, characterized in that: Include the following steps, and follow these steps in sequence: Step 1: Input the field scene image into a lightweight target detection network for multi-scale feature extraction to obtain shallow features, mid-level features and deep semantic features, and simultaneously form a mid-to-high level semantic feature extraction branch and a small target sensitive feature extraction branch. Step 2: Input the deep semantic features obtained in Step 1 into the Foreground Saliency Response Generation Module (FSRG) to generate a foreground saliency response map and calculate the saliency response intensity, which is used to characterize the overall saliency level of potential target regions in the image. Step 3: Input the saliency response map obtained in Step 2 and the small target sensitive features obtained in Step 1 into the saliency-guided feature enhancement module SDEM. Adaptively select the saliency-guided enhancement path or the weak enhancement / residual preservation path according to the saliency response intensity, and output the enhanced features in combination with the local detail branch. Step 4: Input the enhanced features obtained in Step 3 into the multi-scale feature fusion module, and concatenate and integrate them with the deep semantic features to obtain the fused features; Step 5: Input the fused features obtained in Step 4 into the detection and prediction module to obtain the target category probability and bounding box position parameters, and complete the detection of small targets in the field environment; Step 6: Train the entire network end-to-end using a detection task-driven approach to optimize the foreground saliency response generation process, the saliency-guided feature enhancement process, and the target detection process in a coordinated manner.

2. The lightweight detection method for small targets in the field based on endogenous saliency guidance according to claim 1, characterized in that: The lightweight object detection network in step one extracts features from the input image through convolutional layers and layer-by-layer downsampling operations to obtain a multi-scale feature set. ,in These are shallow, high-resolution features. This is a mid-level feature. For deep semantic features; the mid-to-high-level semantic feature extraction branch uses deep semantic features As input for foreground saliency response modeling, the small target sensitive feature extraction branch retains high-resolution local structural features for subsequent saliency-guided enhancement.

3. The lightweight detection method for small targets in the field based on endogenous saliency guidance according to claim 1, characterized in that: The foreground saliency response generation module FSRG in step two generates a saliency response map by performing convolutional mapping, nonlinear activation, and channel compression operations on deep semantic features. The process is represented as follows: in, Used to extract local spatial context information Used to implement channel compression and feature mapping This represents the Sigmoid activation function; further, statistical analysis of the significance response plot yields the significance response intensity: in, and These represent the height and width of the saliency response map, respectively.

4. The lightweight detection method for small targets in the field based on endogenous saliency guidance according to claim 1, characterized in that: The saliency-guided feature enhancement module (SDEM) in step three includes a saliency modulation branch, a local detail enhancement branch, and a residual preservation branch; wherein, firstly, the saliency response map is sized and aligned, denoted as... Then, element-wise multiplication is used to apply the saliency response map to the small target sensitive features to generate foreground candidate features. The process is as follows: in, Based on enhanced features, This indicates element-wise multiplication.

5. A lightweight detection method for small targets in the field based on endogenous saliency guidance according to claim 1, characterized in that: In step three, the method of adaptively selecting the saliency-guided enhancement path or the weak enhancement / residual preservation path based on the saliency response intensity is as follows: Let the enhancement threshold be... ,when When entering the saliency-guided enhancement path, its output is represented as: when When entering a weak enhancement or residual preservation path, its output is expressed as: in, To enhance the coefficient, It is a weak enhancement coefficient, and Simultaneously, the local detail enhancement branch performs local convolution modeling on the basic enhancement features to obtain local detail features: The final enhanced features are represented as follows:

6. A lightweight detection method for small targets in the field based on endogenous saliency guidance as described in claim 1, characterized in that: The multi-scale feature fusion and detection prediction process in steps four and five is as follows: enhancing features... With deep semantic features By splicing the components together, the fused features are obtained: Subsequently passed Convolution performs channel compression and feature integration to obtain the final fused features: The fused features are then input into the detection and prediction module, where the classification branch outputs the target category probability and the regression branch outputs the target bounding box parameters. Simultaneously, the classification loss function and the bounding box regression loss function are used to jointly optimize the entire network, and its overall loss function is expressed as:

Images