A small target detection model based on multi-scale feature fusion

Abstract

This invention provides a small target detection model based on multi-scale feature fusion, relating to the technical field of detection models. The model includes: a backbone network comprising an initial downsampling module, a complementary feature downsampling module, a C2PSA module, and an SPPF module connected sequentially; the complementary feature downsampling module mitigates feature information loss during downsampling; a neck network comprising a cross-level feature concatenation module and a multi-level feature enhancement module; the cross-level feature concatenation module employs a dimensionality reconstruction upsampling strategy to achieve multi-scale feature fusion; the multi-level feature enhancement module improves feature representation through a spatial weight enhancement mechanism; and a detection head network comprising detection heads at three resolution scales: 160×160, 80×80, and 40×40, for multi-scale target detection. This invention enhances the preservation of target details through dual-path feature preservation and spatial-semantic information complementarity.

Description

Technical Field

[0001] This invention relates to the technical field of detection models, and more particularly to a small target detection model based on multi-scale feature fusion. Background Technology

[0002] Existing research mainly optimizes small target detection performance from two directions: improving accuracy and advancing lightweight design. However, models that prioritize accuracy often come with a significant increase in parameters and computational costs, leading to deployment difficulties; while lightweight models struggle to achieve high detection accuracy simultaneously. Therefore, further improving the accuracy of small target detection while controlling computational resource consumption remains a key challenge that urgently needs to be addressed. Summary of the Invention

[0003] In response to the technical problems mentioned in the background section, this invention provides a small target detection model based on multi-scale feature fusion. This invention effectively alleviates the problems of feature loss and inaccurate localization of small targets, improving detection accuracy and maintaining strong detection capabilities even at high confidence levels. It also provides a feasible solution for real-time target detection in resource-constrained environments.

[0004] The technical means employed in this invention are as follows:

[0005] A small target detection model based on multi-scale feature fusion includes: The backbone network includes: an initial downsampling module, a feature complementary downsampling module, a C2PSA module, and an SPPF module connected in sequence; the feature complementary downsampling module is used to mitigate the loss of feature information during the downsampling process. The neck network includes: a cross-level feature concatenation module and a multi-level feature enhancement module; the cross-level feature concatenation module adopts a dimensionality reconstruction upsampling strategy to achieve multi-scale feature fusion; the multi-level feature enhancement module improves feature representation capability through a spatial weight enhancement mechanism; The detection head network includes detection heads at three resolution scales: 160×160, 80×80, and 40×40, for multi-scale target detection.

[0006] Furthermore, the feature complementary downsampling module includes: The channel compression unit is used to compress the input features in terms of channel dimensions. The first downsampling branch processes the compressed features using traditional downsampling methods; The second downsampling branch uses pooling and convolution fusion downsampling to process the compressed features; The feature concatenation unit is used to concatenate the output features of two branches; The convolutional unit uses 3×3 convolution and 1×1 convolution to perform cross-path fusion and channel adjustment on the concatenated features in sequence.

[0007] Furthermore, the pooling-convolutional fusion downsampling module includes: The pooling downsampling branch uses a 2×2 pooling kernel to perform a pooling operation with a step size of 2; The pooling convolution downsampling branch includes a first sub-branch and a second sub-branch. The first sub-branch uses an R1×R1 convolution kernel to extract detailed features, and the second sub-branch uses an R1×R1 pooling kernel to generate spatially distributed weights. , Indicates the receptive field parameters; The element-wise addition unit is used to add the outputs of two sub-branches element by element; The downsampling convolutional unit uses an R2×R2 convolutional kernel for downsampling, where R2=Rf-R1+1.

[0008] Furthermore, the cross-level feature splicing module includes: The feature partitioning unit is used to divide the shallow input features into a first sub-feature and a second sub-feature in the channel dimension; The dimension reconstruction upsampling unit is used to expand the resolution of deep input features. The expansion is achieved through channel dimension expansion and sub-pixel recombination operations. The first fusion unit uses a 2×2 convolution to fuse the expanded deep features with the second sub-feature; The spatial alignment unit uses 2×2 convolution to align the spatial dimensions of the expanded deep features. The second fusion unit uses a 2×2 convolution to concatenate and fuse the aligned features with the output of the first fusion unit; The final splicing unit is used to splice the first sub-feature with the output of the second fusion unit, thereby realizing the effective transmission of low-level location information to high-level features.

[0009] Furthermore, the dimension reconstruction upsampling unit controls the resolution enhancement by using an expansion factor r. After expanding the channel dimension of the input feature map by a multiple of r×r, it achieves the mapping from channel to spatial dimension through sub-pixel recombination operation.

[0010] Furthermore, the multi-level feature enhancement module includes: The channel adjustment unit uses a 1×1 convolution to adjust the channels of the input features; The first processing branch uses conventional convolutional layers for feature extraction; The second processing branch uses a target enhancement module to generate a spatial weight map. The target enhancement module includes a max pooling layer and an average pooling layer, which are used to extract local salient features and generate boundary smoothing features, respectively. The weight fusion unit is used to perform element-wise multiplication of the spatial weight map with the output of the first processing branch; Cascaded processing units, consisting of cascaded 2×2 convolutional layers, are used to progressively enhance features.

[0011] Furthermore, the target enhancement module includes: The channel equalization unit is used to divide the input features into two groups equally along the channel dimension; The max pooling branch uses a max pooling layer to extract the most salient features within a local region, generating location-enhanced features. The average pooling branch uses an average pooling layer to generate boundary smoothing features, and the average pooling layer has a larger receptive field than the max pooling layer. The feature addition unit is used to add the location enhancement feature and the boundary smoothing feature to form a composite feature map; The convolutional compression unit uses 2×2 convolution to compress the composite feature map into a single channel, generating a spatial weight map that highlights the key target region.

[0012] Furthermore, in the feature pyramid structure of the neck network, the feature map of layer P4 is concatenated with layers P3 and P5 respectively. Layer P5, as a shallow feature, is fused with layer P4, which has a similar scale, to introduce high-level semantic information and control computational complexity.

[0013] Furthermore, the 160×160 high-resolution detector in the detector network is used to enhance the perception of small targets, while the 20×20 large-scale detector is removed.

[0014] Furthermore, in the pooling convolution downsampling branch, the convolution branch uses an R1×R1 convolution kernel, and the pooling branch uses an R1×R1 pooling kernel of the same size. By adding elements one by one, the feature semantic information and spatial distribution information are complemented.

[0015] Compared with the prior art, the present invention has the following advantages: This invention enhances the ability to preserve target details by employing a dual-path feature preservation and spatial-semantic information complementarity method. In the neck network, an upsampling and cross-layer splicing mechanism based on channel-dimensional reconstruction expands the receptive field while preserving target boundary information. At the neck terminal, multiple receptive field features and a spatial weight enhancement mechanism are fused to improve adaptability to complex occlusion and dense target scenes. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the CCA-YOLO model proposed by the present invention based on the YOLOv11 architecture.

[0018] Figure 2 This is a schematic diagram of the feature complementary downsampling module proposed in this invention.

[0019] Figure 3 A schematic diagram of the cross-level feature splicing module designed for this invention is shown.

[0020] Figure 4 This is a schematic diagram of the Expand module of the present invention.

[0021] Figure 5 This is a schematic diagram of the multi-level feature enhancement module of the present invention.

[0022] Figure 6 This is a schematic diagram comparing the detection performance of YOLOv11s and CCA-YOLO of the present invention (Figure I).

[0023] Figure 7 This is Schematic diagram II comparing the detection performance of YOLOv11s and CCA-YOLO of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] like Figure 1-7 As shown, this invention provides a small target detection model based on multi-scale feature enhancement, comprising three main parts: a backbone network, a neck network, and a head.

[0027] In the backbone network structure, the input image first undergoes an initial downsampling operation through the CBS module 101. The CBS module 101 consists of a convolutional layer, a batch normalization layer, and a SiLU activation function connected in series. In this embodiment, the input image size is 640×640 pixels, and after passing through the CBS module 101, the feature map size is reduced to 320×320.

[0028] After initial downsampling, the feature map enters the Complementary Feature Desampling (CFD) module 102 for deep feature extraction. The CFD module 102 adopts a differentiated downsampling compensation mechanism, which effectively alleviates the feature loss problem during the downsampling process through the parallel structure of the traditional downsampling path and the pooling convolution fusion downsampling (PCID) path.

[0029] At the end of the backbone network, C2PSA module 103 and SPPF module 104 are connected sequentially. C2PSA module 103 realizes cross-stage feature optimization, and SPPF module 104 realizes multi-scale context fusion, outputting deep features rich in semantic information.

[0030] The neck network employs a cross-level feature concatenation module (CLFC)105 to construct a feature pyramid. The CLFC module105 uses an upsampling strategy based on tensor dimension reconstruction instead of the traditional nearest neighbor upsampling, which reduces computational overhead while enhancing the ability to extract edge features of small targets.

[0031] The Spatial Feature Enhancement Module (MFE) 106, as a core component of the neck network, performs semantic enhancement on features based on the location and information preservation already completed in the previous stages of the network, thus completing the feature extraction task in the final stage of the feature pyramid. The specific structure of the MFE module 106 will be described in detail in Example 4.

[0032] This embodiment employs three detection heads with resolutions of 160×160, 80×80, and 40×40 (107, 108, and 109). Addressing the challenge of detecting small targets in aerial photography scenarios, this embodiment adds a high-resolution 160×160 detection head (P5) to the existing 80×80 and 40×40 detection heads, while removing the traditional 20×20 resolution large-scale detection head to enhance the perception capability for small targets.

[0033] like Figure 2 As shown, the Complementary Feature Downsampling (CFD) module includes a channel compression unit, a traditional downsampling branch, a PCID branch, a feature stitching unit, and a channel adjustment unit.

[0034] The input feature map X first undergoes channel compression through a 1×1 convolutional layer, reducing the number of channels from... Reduced to This reduces the complexity and number of parameters in subsequent calculations. In this embodiment, =256, =128.

[0035] The compressed features are fed into the traditional downsampling branch and the PCID branch, respectively.

[0036] The traditional downsampling branch uses a 3×3 convolution kernel with a stride of 2 and padding of 1 to achieve standard downsampling operation, halving the size of the output feature map while maintaining 128 channels.

[0037] The PCID branch employs a pooling convolution fusion mechanism, specifically including: Pooling downsampling sub-branch a: using a 2×2 max pooling kernel with a stride of 2. The 2×2 pooling kernel is chosen instead of a 3×3 pooling kernel because when the stride is 2, 3×3 pooling would cause a single output position to correspond to a 5×5 region in the input feature map, resulting in local salient features masking the discriminative features of surrounding small targets.

[0038] PCD sub-branch b: Employs a pooling convolutional downsampling structure. The input features are first divided into two sub-branches: Convolutional branch: R1×R1 convolution kernels are used to extract detailed features, where , In this embodiment, the receptive field parameters are... =3, therefore =2; Pooling branch: Spatially distributed weights are generated using a 2×2 max pooling kernel; The outputs of the two branches are element-wise added to achieve complementarity between feature semantic information and spatial distribution information. The fused features are then downsampled through an R2×R2 convolution, where... The receptive field of this 2×2 convolution is consistent with that of a traditional downsampling layer, ensuring spatial alignment between multiple paths during subsequent feature fusion.

[0039] The mathematical expression for the PCID branch is: ; in, This represents a max pooling operation with a step size of 1 and a pooling kernel size of R1. This indicates an element-wise addition operation. Represents the SiLU function. This represents the convolution kernel used for feature extraction. This represents the bias value used for the convolution in feature extraction. This represents the convolution operation. This represents the convolution weights used for downsampling. This represents the bias value used for downsampling.

[0040] The outputs of the PCID branch and the traditional downsampling branch are concatenated in the feature concatenation unit to achieve channel-level concatenation. The concatenated feature has 256 channels. Subsequently, the concatenated features are fused across paths via 3×3 convolutions and the channel number is adjusted via 1×1 convolutions. The final output feature map with C_out=256 channels is then fed into subsequent network layers.

[0041] The Cross-Level Feature Concatenation Module (CLFC) is used for multi-scale feature fusion in the neck network, including a feature segmentation unit, an expansion unit, a first fusion unit, a spatial alignment unit, a second fusion unit, and a final concatenation unit.

[0042] The CLFC module receives two inputs: a deep low-resolution feature Y (size H×W×C_deep) and a shallow high-resolution feature X (size 2H×2W×C_shallow). In this embodiment, Y is an 80×80×512 feature map from the deep network, and X is a 160×160×256 feature map from the shallow network.

[0043] The shallow input feature X is divided into two independent branches along the channel dimension in the feature segmentation unit: Branch A: Preserves precise location details, with a channel count of [number missing]. / 2=128; Branch b: Participates in cross-layer feature fusion to enrich semantic information, with a channel count of [number missing]. / 2=128; The splitting operation is represented as:

[0044] The resolution of the deep input feature Y is improved through an expansion unit. The expansion unit employs a channel-dimensional expansion and recombination design, with the expansion factor r controlling the magnitude of the resolution increase. In this embodiment, r=2, meaning the channel dimension of the input feature map is expanded by a factor of 2×2=4 (from 512 to 8). Subsequently, a pixel shaving operation is used to map the channels to the spatial dimension, resulting in an output feature map size of 160×160×512.

[0045] The extended operation is represented as:

[0046] The first fusion unit will expand the deep features The shallow features from branch b are fused together using a 2×2 convolution operation:

[0047] This fusion operation enables the rich details such as edges and textures in low-level features to supplement the spatial accuracy lost by high-level features during multiple downsampling processes, thereby strengthening the interaction between shallow and deep features.

[0048] Spatial alignment units for extended deep features Perform a single 2×2 convolution operation to align the spatial dimensions of features across layers:

[0049] The second fusion unit concatenates two spatially aligned feature maps y_1 and y_2 and then performs a 2×2 convolution to achieve deep fusion of cross-layer information:

[0050] The final stitching unit stitches the enhanced deep feature y' with the branch a that retains rich spatial details, thus achieving effective transfer of low-level location information to high-level features: ; The final output feature map z has a size of 160×160×640, which includes both the depth features of the target and preserves shallow features and contextual information. The multi-level feature enhancement module (MFE) adopts a progressive enhancement mode, including an input adjustment unit, a multi-level feature extraction unit, a target enhancement module (TEM), and a weight fusion module (WFM). The input feature map is first processed through a 1×1 convolutional layer to adjust the number of channels, uniformly adjusting the number of channels to C_mid=256.

[0051] The main path uses 2×2 convolutional layers 402 and 403 as the core units for feature extraction at each level. Using two cascaded 2×2 convolutional layers instead of a single 3×3 convolutional layer achieves the same receptive field as a 3×3 convolutional layer while providing stronger detail enhancement and feature preservation capabilities during hierarchical enhancement.

[0052] The Target Enhancement Module (TEM) works in parallel with the main path. Taking the first-level TEM 404 as an example, its structure includes: Channel Segmentation: The input features are divided into two groups along the channel dimension, with 128 channels in each group; Location Enhancement Branch: A 2×2 max pooling layer is used to extract the most salient features in local regions to generate location enhancement features, amplifying areas with weak responses due to small size; Boundary Smoothing Branch: A 3×3 average pooling layer (with a larger receptive field) is used to generate boundary smoothing features, utilizing the smoothing property to make the response value of the feature map gradually decrease from the center of the target to the edge; Feature Fusion: The location enhancement features and boundary smoothing features are added element-wise to form a composite feature map, making the response stronger closer to the center of the target and weaker closer to the edge; Weight Generation: A 2×2 convolutional layer compresses the composite feature map into a single channel to generate a spatial weight map that highlights key target regions. In the second-level TEM 405, the max pooling layer uses a 2×2 pooling kernel, and the feature compression convolution also uses a 2×2 convolutional kernel. Since the overall receptive field of TEM is 3×3, two convolutions are used to align the outputs of the main path and the TEM path at the semantic level.

[0053] The weighted fusion modules (WFM) 406 and 407 expand the single-channel spatial weight map to match the number of channels (256 channels) of the convolutional feature map through channel copying, and then add it element-wise with the backbone path feature map. Since the spatial weight map enhances positional and boundary information, fusion can effectively improve the spatial perception and feature discrimination of the backbone path feature map.

[0054] After two stages of feature enhancement and extraction, the final output is a 1×1 convolutional layer with 408 channels and 256 output channels. This dense and multi-stage enhancement strategy enables the model to construct a strong feature response even when the target features are sparse or densely occluded.

[0055] This embodiment optimizes the neck network structure by employing a cross-level feature concatenation strategy. The feature map of layer P4 (80×80 resolution) is concatenated with layers P3 (160×160) and P5 (40×40), respectively. This design is based on the following considerations: the positional information of layer P4 is clear and does not require reliance on lower-level features to complete the position, so it is directly upsampled and fused with layer P3; layer P5, as a deep feature, needs to incorporate high-level semantics to enhance its expressive power, therefore layer P5 is fused with layer P4, which has a similar scale, rather than directly with layer P3; this design avoids feature loss and background noise interference caused by excessive scale difference (4-fold difference) between P5 and P3, effectively controlling computational complexity while maintaining feature expressive power.

[0056] The features of layer P5 are upsampled by 2 times by the CLFC module and fused with the features of layer P4 to output P4'; P4' is upsampled by 2 times by the CLFC module and fused with the features of layer P3 to output P3'; P3' is enhanced by the MFE module and then sent to a 160×160 detector head; P4' is enhanced by the MFE module and then sent to an 80×80 detector head; the features of layer P5 are enhanced by the MFE module and then sent to a 40×40 detector head.

[0057] This embodiment applies the above-mentioned small target detection model to the task of small target detection in aerial images. Aerial images typically have the following characteristics: (1) high shooting altitude, resulting in targets appearing extremely small in the image; (2) complex background, containing interfering information such as buildings, vegetation, and roads; (3) dense target distribution, with occlusion phenomena. The 20×20 resolution detector in the traditional detection model is difficult to effectively detect small targets, while the 80×80 resolution detector also has difficulty recognizing extremely small targets (less than 16×16 pixels).

[0058] In this embodiment, the model input image size is 640×640 pixels, the batch size is 16, and the number of training epochs is [number missing]. The optimizer uses SGD with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005. The learning rate scheduling uses a cosine annealing strategy.

[0059] 160×160 detection head: responsible for detecting extremely small targets with a size ranging from 8×8 to 32×32 pixels, with the anchor frame size set to (10,13), (16,30), (33,23); 80×80 detection head: responsible for detecting small targets with a size ranging from 32×32 to 96×96 pixels, with the anchor frame size set to (30,61), (62,45), (59,119); 40×40 detection head: responsible for detecting medium-sized targets with a size ranging from 96×96 to 320×320 pixels, with anchor frame sizes set to (116,90), (156,198), and (373,326).

[0060] Test results on the VisDrone9 aerial dataset show that, compared to the baseline model (YOLOv8n), the model in this embodiment improves the average precision (AP) by 8.5% for small target categories (such as pedestrians and vehicles) and the recall rate by 12.3% for very small targets (less than 16×16 pixels), while the number of model parameters only increases by 5.2% and the computational cost increases by 7.8%, achieving a good balance between accuracy and efficiency.

[0061] Example 1

[0062] To verify the effectiveness of the various improvements in this invention, ablation experiments were conducted on the VisDrone9 dataset. The experiments used precision, recall, and mAP as metrics. 50 mAP 75 mAP 50-95 The number of parameters and the amount of computation are used as unified evaluation indicators for all categories.

[0063] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the above embodiments of the present invention, the descriptions of each embodiment have their own emphasis; parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. It should be understood that the disclosed technical content in the several embodiments provided in this application can be implemented in other ways.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A small target detection model based on multi-scale feature fusion, characterized in that, include: The backbone network includes: an initial downsampling module, a feature complementary downsampling module, a C2PSA module, and an SPPF module connected in sequence; the feature complementary downsampling module is used to mitigate the loss of feature information during the downsampling process. The neck network includes: a cross-level feature concatenation module and a multi-level feature enhancement module; the cross-level feature concatenation module adopts a dimensionality reconstruction upsampling strategy to achieve multi-scale feature fusion; the multi-level feature enhancement module improves feature representation capability through a spatial weight enhancement mechanism; The detection head network includes detection heads at three resolution scales: 160×160, 80×80, and 40×40, for multi-scale target detection.

2. The small target detection model based on multi-scale feature fusion according to claim 1, characterized in that, The feature complementary downsampling module includes: The channel compression unit is used to compress the input features in terms of channel dimensions. The first downsampling branch processes the compressed features using traditional downsampling methods; The second downsampling branch uses pooling and convolution fusion downsampling to process the compressed features; The feature concatenation unit is used to concatenate the output features of two branches; The convolutional unit uses 3×3 convolution and 1×1 convolution to perform cross-path fusion and channel adjustment on the concatenated features in sequence.

3. The small target detection model based on multi-scale feature fusion according to claim 2, characterized in that, The pooling-convolutional fusion downsampling module includes: The pooling downsampling branch uses a 2×2 pooling kernel to perform a pooling operation with a step size of 2; The pooling convolution downsampling branch includes a first sub-branch and a second sub-branch. The first sub-branch uses an R1×R1 convolution kernel to extract detailed features, and the second sub-branch uses an R1×R1 pooling kernel to generate spatially distributed weights. , Indicates the receptive field parameters; The element-wise addition unit is used to add the outputs of two sub-branches element by element; The downsampling convolutional unit uses an R2×R2 convolutional kernel for downsampling, where R2=Rf-R1+1.

4. The small target detection model based on multi-scale feature fusion according to claim 1, characterized in that, The cross-level feature splicing module includes: The feature partitioning unit is used to divide the shallow input features into a first sub-feature and a second sub-feature in the channel dimension; The dimension reconstruction upsampling unit is used to expand the resolution of deep input features. The expansion is achieved through channel dimension expansion and sub-pixel recombination operations. The first fusion unit uses a 2×2 convolution to fuse the expanded deep features with the second sub-feature; The spatial alignment unit uses 2×2 convolution to align the spatial dimensions of the expanded deep features. The second fusion unit uses a 2×2 convolution to concatenate and fuse the aligned features with the output of the first fusion unit; The final splicing unit is used to splice the first sub-feature with the output of the second fusion unit, thereby realizing the effective transmission of low-level location information to high-level features.

5. The small target detection model based on multi-scale feature fusion according to claim 4, characterized in that, The dimension reconstruction upsampling unit controls the resolution enhancement by using an expansion factor r. After expanding the channel dimension of the input feature map by a multiple of r×r, it achieves the mapping from channel to spatial dimension through sub-pixel recombination operation.

6. The small target detection model based on multi-scale feature fusion according to claim 1, characterized in that, The multi-level feature enhancement module includes: The channel adjustment unit uses a 1×1 convolution to adjust the channels of the input features; The first processing branch uses conventional convolutional layers for feature extraction; The second processing branch uses a target enhancement module to generate a spatial weight map. The target enhancement module includes a max pooling layer and an average pooling layer, which are used to extract local salient features and generate boundary smoothing features, respectively. The weight fusion unit is used to perform element-wise multiplication of the spatial weight map with the output of the first processing branch; Cascaded processing units, consisting of cascaded 2×2 convolutional layers, are used to progressively enhance features.

7. The small target detection model based on multi-scale feature fusion according to claim 6, characterized in that, The target enhancement module includes: The channel equalization unit is used to divide the input features into two groups equally along the channel dimension; The max pooling branch uses a max pooling layer to extract the most salient features within a local region, generating location-enhanced features. The average pooling branch uses an average pooling layer to generate boundary smoothing features, and the average pooling layer has a larger receptive field than the max pooling layer. The feature addition unit is used to add the location enhancement feature and the boundary smoothing feature to form a composite feature map; The convolutional compression unit uses 2×2 convolution to compress the composite feature map into a single channel, generating a spatial weight map that highlights the key target region.

8. The small target detection model based on multi-scale feature fusion according to claim 1, characterized in that, In the feature pyramid structure of the neck network, the feature map of layer P4 is concatenated with layers P3 and P5 respectively. Layer P5, as a shallow feature, is fused with layer P4, which has a similar scale, to introduce high-level semantic information and control computational complexity.

9. The small target detection model based on multi-scale feature fusion according to any one of claims 1-8, characterized in that, The 160×160 high-resolution detector in the detector network is used to enhance the perception of small targets, while the 20×20 large-scale detector is removed.

10. The small target detection model based on multi-scale feature fusion according to claim 2, characterized in that, In the pooling convolution downsampling branch, the convolution branch uses an R1×R1 convolution kernel, and the pooling branch uses an R1×R1 pooling kernel of the same size. The feature semantic information and spatial distribution information are complemented by adding elements one by one.

Images