Small target detection method and device based on neural network, server and storage medium

Abstract

The application discloses a small target detection method and device based on a neural network, a server and a storage medium, and belongs to the technical field of target detection. The method comprises the following steps: processing primary features through a convolution and an MDPEM module to obtain high-resolution detailed features; processing the high-resolution detailed features through the convolution and the MDPEM module to obtain medium-resolution balanced features; processing the medium-resolution balanced features through a multi-scale wavelet downsampling layer and the MDPEM module to obtain low-resolution semantic features; processing the low-resolution semantic features through an AIFI to obtain attention-weighted semantic features; inputting the attention-weighted semantic features, the medium-resolution balanced features and the high-resolution detailed features into a fusion architecture to obtain deep fusion features; and inputting the deep fusion features into a decoder to obtain a small target detection result. Through multi-scale feature enhancement and the fusion architecture, accurate detection of small targets is realized.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method, apparatus, server, and storage medium for small target detection based on neural networks. Background Technology

[0002] With the popularization of UAV remote sensing technology, dynamic monitoring based on aerial imagery is playing an increasingly important role in fields such as security and disaster early warning. Among these, the accurate detection of tiny targets in the images is the core link to ensure the monitoring effect, and it is also one of the current research challenges.

[0003] Existing detection methods mainly fall into two categories: one relies on manual interpretation based on operator experience, which, while applicable in specific scenarios, is inefficient and prone to false negatives. The other is based on detection architectures using traditional convolutional neural networks, such as the YOLO series and Faster R-CNN models. These methods typically extract features through a backbone network and use feature pyramids for multi-scale fusion to perform object detection.

[0004] However, the above methods have significant limitations when dealing with small targets. Due to the small target size, the ability to represent the features of small targets is insufficient, ultimately leading to missed detections. Summary of the Invention

[0005] This invention provides a method, apparatus, server, and storage medium for small target detection based on neural networks, in order to solve the technical problem of low accuracy in small target detection caused by insufficient feature extraction capabilities in the prior art.

[0006] In a first aspect, embodiments of the present invention provide a small target detection method based on a neural network, comprising:

[0007] The primary features extracted from the image are processed by convolution and MDPEM modules to enhance the details of small targets and obtain high-resolution detail features.

[0008] The high-resolution detail features are processed by convolution and MDPEM modules to balance detail and semantic information, resulting in medium-resolution balanced features.

[0009] The medium-resolution balanced features are processed by a multi-scale wavelet downsampling layer and an MDPEM module to extract deep semantics of small targets and reduce dimensionality, resulting in low-resolution semantic features.

[0010] The MDPEM module includes a three-layer parallel branch structure consisting of a multi-scale depthwise separable convolutional layer, a deep convolutional layer, and an integrated EPGO cue guidance mechanism. It is used to extract multi-granularity spatial context features and achieve multi-scale enhancement and dynamic perception of small target features by adjusting feature resolution and information density.

[0011] The low-resolution semantic features are processed by AIFI to obtain attention-weighted semantic features;

[0012] The attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features are sequentially input into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features, which are used to integrate semantic information at different resolutions, and the adaptive reparameterization feature enhancement layer is used to enhance the semantic features of small targets.

[0013] The deep fusion features are input into the decoder to obtain the small target detection results.

[0014] Secondly, embodiments of the present invention also provide a small target detection device based on a neural network, comprising:

[0015] The high-resolution detail feature acquisition module is used to process the primary features extracted from the image through convolution and MDPEM modules to enhance the details of small targets and obtain high-resolution detail features.

[0016] The medium-resolution balanced feature acquisition module is used to process the high-resolution detail features through convolution and MDPEM modules to balance details and semantic information, thereby obtaining medium-resolution balanced features.

[0017] The low-resolution semantic feature acquisition module is used to process the medium-resolution balanced features through a multi-scale wavelet downsampling layer and an MDPEM module to extract the deep semantics of small targets and reduce their dimensionality, thereby obtaining low-resolution semantic features.

[0018] The attention-weighted module is used to process the low-resolution semantic features through AIFI to obtain attention-weighted semantic features;

[0019] The feature fusion module is used to sequentially input the attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features, which are used to integrate semantic information at different resolutions and enhance the semantic features of small targets using the adaptive reparameterization feature enhancement layer.

[0020] The target detection module is used to input the deep fusion features into the decoder to obtain small target detection results.

[0021] Thirdly, embodiments of the present invention also provide a server, comprising:

[0022] One or more processors;

[0023] Storage device for storing one or more programs;

[0024] When the one or more programs are executed by the one or more processors, the one or more processors implement the neural network-based small target detection method provided in the above embodiments.

[0025] Fourthly, embodiments of the present invention also provide a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the small target detection method based on a neural network as provided in the above embodiments.

[0026] The small target detection method, device, server, and storage medium based on neural networks provided in this invention process the primary features extracted from the image through convolution and MDPEM modules to enhance the details of small targets, obtaining high-resolution detail features. The high-resolution detail features are then processed through convolution and MDPEM modules to balance detail and semantic information, resulting in medium-resolution balanced features. These medium-resolution balanced features are then processed through a multi-scale wavelet downsampling layer and an MDPEM module to extract deep semantics of the small targets and reduce their dimensionality, yielding low-resolution semantic features. The MDPEM module includes a multi-scale depth-separable convolutional layer, a depth convolutional layer, and a three-layer parallel branch integrating an EPGO cue guidance mechanism. A structure is used to extract multi-granularity spatial context features and, by adjusting feature resolution and information density, achieve multi-scale enhancement and dynamic perception of small target features. The low-resolution semantic features are processed by AIFI to obtain attention-weighted semantic features. These attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features are sequentially input into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features. These features integrate semantic information at different resolutions, and the adaptive reparameterization feature enhancement layer further enhances the semantic features of small targets. The deep fusion features are then input into a decoder to obtain the small target detection result. Through this multi-scale feature enhancement and fusion architecture, features at different resolutions, from detail to semantics, are extracted and strengthened, enabling the perception and integration of small target features and ultimately improving the accuracy of small target detection. Attached Figure Description

[0027] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0028] Figure 1 This is a flowchart of the small target detection method based on neural networks provided in Embodiment 1 of the present invention;

[0029] Figure 2This is a schematic diagram of the structure of the MDPEM module in the small target detection method based on neural networks provided in Embodiment 1 of the present invention;

[0030] Figure 3 This is a schematic diagram of the structure of the multi-scale wavelet downsampling layer in the small target detection method based on neural networks provided in Embodiment 1 of the present invention;

[0031] Figure 4 This is a schematic diagram of the structure of the multi-scale wavelet upsampling layer in the small target detection method based on neural networks provided in Embodiment 1 of the present invention;

[0032] Figure 5 This is a schematic diagram of the adaptive reparameterized feature enhancement layer in the small target detection method based on neural networks provided in Embodiment 1 of the present invention;

[0033] Figure 6 This is a structural diagram of the small target detection device based on a neural network provided in Embodiment 2 of the present invention;

[0034] Figure 7 This is a structural diagram of the server provided in Embodiment 3 of the present invention. Detailed Implementation

[0035] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0036] Example 1

[0037] Figure 1 This is a flowchart of a small target detection method based on a neural network provided in Embodiment 1 of the present invention. This embodiment is applicable to the detection of small targets in images of various scenes, and specifically includes the following steps:

[0038] Step 110 involves processing the primary features extracted from the image through convolution and MDPEM modules to enhance the details of small targets, thereby obtaining high-resolution detail features.

[0039] An image refers to pre-processed small target image data input to a neural network model. Primary features refer to the features obtained after the image is input into the network model and processed through a convolutional layer.

[0040] For example, for small target detection by drones, images of small targets are acquired by drones equipped with high-resolution visible light or infrared sensors. After preprocessing, the acquired images are input into a network model. First, convolution is performed on the images to extract primary features. Then, the extracted primary features are input into the convolution and MDPEM modules to enhance the details of the small targets, and high-resolution detail features are output.

[0041] For example, such as Figure 2 As shown, the MDPEM module comprises a three-layer parallel branch structure consisting of a multi-scale depthwise separable convolutional layer, a depthwise convolutional layer, and an integrated EPGO cue guidance mechanism. The multi-scale depthwise separable convolutional layer includes parallel branches with different receptive fields to extract multi-granular spatial context information for small objects; and convolutional branches for channel-dimensional integration and dimensionality reduction of features. Each parallel branch with different receptive fields contains a depthwise separable convolution (DWConv), with receptive field sizes of 1x1, 3x3, 5x5, and 7x7, respectively. Average pooling is performed after each DWConv to extract multi-granular spatial context information from the input features, thus covering and considering local details. The convolutional branches primarily undergo 1×1 convolution processing. The multi-scale depthwise separable convolutional layer also includes an original feature input branch. The outputs of the parallel branches with different receptive fields, the convolutional branches, and the original feature input branch are then integrated along the channel dimension to obtain multi-scale integrated features. The obtained multi-scale integrated features are then refined through deep convolutional layers. The deep convolutional layers perform a first deep separable convolution on the output features of the multi-scale deep separable convolutional layer to obtain a first intermediate feature; the first intermediate feature is then subjected to average pooling and a second deep separable convolution to obtain a second intermediate feature; the output features of the multi-scale deep separable convolutional layer, the first intermediate feature, and the second intermediate feature are summed by channel dimension to obtain features adjusted for spatial resolution and information density. The first deep separable convolution can be a 3×3 DWConv, and the second deep separable convolution can be a 1×1 DWConv. Specifically, the multi-scale integrated features are processed by a 3×3 DWConv to obtain primary refined features, and the primary refined features are then processed by average pooling and a 1×1 DWConv to obtain secondary refined features. Finally, the multi-scale integrated features, primary refined features, and secondary refined features are summed by channel dimension to adjust the spatial resolution and information density of the feature map, resulting in features with adjusted spatial resolution and information density, ensuring that subsequent feature processing is performed at an appropriate scale. The features with adjusted spatial resolution and information density are then input into a three-layer parallel branch structure integrating the EPGO cue guidance mechanism to obtain multi-scale enhanced features.

[0042] For example, the three-layer parallel branch structure integrating the EPGO prompting mechanism is used to input the spatially and informationally density-adjusted features into the first, second, and third branches respectively. Element-wise dot products are performed on the output features of the second and third branches to obtain the dot-multiplied features. The dot-multiplied features are then summed with the output features of the first branch and the spatially and informationally density-adjusted features to obtain the summed features. The summed features are then convolved to obtain multi-scale enhanced features. The first branch includes an EPGO module, used to extract the global context information of the spatially and informationally density-adjusted features and generate a scaling factor adapted to the channel dimension. The second branch is used to form an information flow from the spatially and informationally density-adjusted features through a linear transformation layer, and then generate a mask between 0 and 1 using ReLU and an activation function to adaptively emphasize important feature regions. The third branch is used to form an information flow from the spatially and informationally density-adjusted features through a linear transformation layer, and then extract the local spatial information of small targets through separable convolution processing, followed by an activation function to enhance the ability to obtain important details of small targets. The model can be structured as follows: the first branch is the Efficient Prompt Guidance Operation (EPGO) branch, the second branch is the gating branch, and the third branch is the Efficient Position Aware branch. The EPGO branch primarily includes the EPGO module, which constructs a dynamic prompt generation path through batch normalization layers, linear transformation layers, and activation functions. It extracts global contextual information from features adjusted for spatial resolution and information density, and generates a dynamic scaling factor adapted to the channel dimension. This factor adaptively adjusts the feature selection threshold, enhancing the model's robustness to complex scenes. The Efficient Position Aware branch forms an information flow through linear transformation layers, capturing pixel-level local spatial information of small targets using a 3x3 DWConv layer. An activation function then enhances its non-linear expressive power, thereby enhancing the ability to capture important details of small targets. The gating branch forms an information flow through linear transformation layers, generating a mask between 0 and 1 using ReLU and Sigmoid activation functions to adaptively emphasize the feature regions of small targets. Specifically, the features adjusted for spatial resolution and information density are input into the EPGO, gating, and Efficient Position Aware branches, respectively. Element-wise multiplication is performed on the output features of the gating and Efficient Position Aware branches to obtain the multiplied features. Next, the features obtained after dot product are summed with the output features of the EPGO branch and the features adjusted for spatial resolution and information density, resulting in summed features. Finally, the summed features are convolved to obtain multi-scale enhanced features.

[0043] The MDPEM module enhances global context awareness by introducing the EPGO branch. Combining multi-branch parallel processing and adaptive mechanisms, it achieves refined perception of multi-scale features, high-precision edge information enhancement, and adaptive spatial modulation capabilities through a unique architecture design.

[0044] Step 120: The high-resolution detail features are processed by convolution and MDPEM modules to balance detail and semantic information, resulting in medium-resolution balanced features.

[0045] High-resolution features are spatially large, containing many pixels. They preserve rich and fine local details in an image (such as object edges, textures, and corners), which is crucial for locating small targets, but each feature contains a lower semantic level. Low-resolution features are spatially smaller, obtained by downsampling high-resolution feature maps. They sacrifice fine local details but contain higher-level, more abstract semantic information.

[0046] For example, the high-resolution detail features obtained in step 110 are input again into the convolution and MDPEM modules for processing, such as... Figure 2 As shown, the MDPEM module includes depthwise separable convolution and average pooling operations, which have downsampling capabilities. When the downsampling operation is applied to a high-resolution feature map, the convolution kernel slides with a specific stride, gradually merging feature information from adjacent regions. This process halves the spatial size of the output feature map and increases the number of channels, resulting in a more compact medium-resolution feature map. Based on this, the module achieves a balance between detailed information and semantic features, ultimately outputting a balanced medium-resolution feature map.

[0047] Step 130: The medium-resolution balanced features are processed by a multi-scale wavelet downsampling layer and an MDPEM module to extract the deep semantics of small targets and reduce their dimensionality, thereby obtaining low-resolution semantic features.

[0048] Haar Fast Wavelet Transform is a highly efficient multi-resolution signal analysis tool. The core idea of ​​wavelet transform is to decompose a signal into approximate components and detail components of different resolutions through simple averaging and differencing operations.

[0049] For example, such as Figure 3As shown, the multi-scale wavelet downsampling layer consists of multiple parallel branches and a wavelet forward propagation module. This layer contains N parallel partial convolution (PConv) branches, each consisting of 3×3 PConv, batch normalization, ReLU activation, 1×1 convolution, batch normalization, and ReLU activation, used to refine spatial details from multiple local perspectives and suppress redundant computation. Simultaneously, a pointwise convolution (PWConv) branch is set, with a structure of 1×1 PWConv, batch normalization, ReLU activation, 1×1 convolution, batch normalization, and ReLU activation, used to reconstruct and integrate small target information in the channel dimension, enhancing inter-channel interaction. This layer further introduces a wavelet forward propagation module (WPFP) integrating Haar fast wavelet transform second-order decomposition. The first-level decomposition decomposes the input feature map into four frequency sub-bands: one low-frequency sub-band, carrying the overall contour and structural information of the image; and three high-frequency sub-bands, capturing edge, texture, and other detailed features respectively. The second-level decomposition further divides the low-frequency subband obtained in the first level into four equal parts, generating four more fine-grained subbands. Overall, this process forms a multi-level representation structure containing seven high-frequency subbands and one extremely low-frequency subband. This multi-scale, progressively refined decomposition strategy helps extract the low-frequency structural features upon which small targets depend, while effectively suppressing high-frequency noise interference, thereby improving the discriminative power of the feature representation.

[0050] For example, the medium-resolution balanced features are further input into a multi-scale wavelet downsampling layer and an MDPEM module for processing. This processing significantly reduces the spatial size of the feature map while enhancing the receptive field and semantic expressive power of each feature point, ultimately outputting low-resolution semantic features rich in high-level semantic information.

[0051] Step 140: The low-resolution semantic features are processed by AIFI to obtain attention-weighted semantic features.

[0052] The Attention-Based Intra-Scale Feature Interaction (AIFI) module is a self-attention mechanism module in the Transformer architecture. The core task of AIFI is to model the long-range dependencies between all feature points within a feature map of the same resolution. Unlike convolutional operations that only focus on local neighborhoods, the self-attention mechanism allows each location in the feature map to interact directly with all other locations, thereby evaluating and enhancing the features of each location based on the global context.

[0053] For example, the AIFI module applies self-attention computation to the low-resolution feature map. By analyzing the global correlation between feature points, the model can highlight semantic information highly relevant to small targets and suppress irrelevant or redundant background information. After processing, the final output retains the original deep semantics, and the key internal features are significantly enhanced by attention-weighted semantic features, providing higher-quality and more discriminative high-level semantic input for subsequent multi-scale feature fusion.

[0054] Step 150: The attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features are sequentially input into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features, which are used to integrate semantic information at different resolutions and enhance the semantic features of small targets using the adaptive reparameterization feature enhancement layer.

[0055] For example, such as Figure 4As shown, the multi-scale wavelet upsampling layer includes linear mapping and repeated expansion paths, linear mapping and channel displacement paths, and parallel units. The linear mapping and repeated expansion paths achieve dimensional adaptation and spatial replication of input features by fusing linear mapping and repeated expansion techniques. The repeated expansion technique is implemented through the `repeat` function, which replicates or expands the spatial dimension or number of channels of the input features to match the size requirements of subsequent feature fusion, facilitating the overlay of information from multiple branches. This process ensures that the input, after transformation by the linear layer, can be smoothly transmitted and aligned spatially or dimensionally with features from other paths. The linear mapping and channel displacement paths achieve cross-channel and cross-spatial feature information interaction and recombination by fusing linear mapping and channel displacement techniques, breaking the limitations of the local receptive field in traditional convolution. The channel displacement technique is implemented through the channel displacement (Shift) mechanism. The Shift mechanism is a lightweight feature operation technique whose core lies in achieving interaction and recombination of information from different channels or spatial locations by cyclically shifting the feature tensor through channels or spatial dimensions. This mechanism does not rely on convolutional computation; it effectively expands the receptive field of features solely through index shifting, promoting the fusion of cross-regional contextual information. While significantly reducing computational complexity, it provides a richer input structure for subsequent feature extraction and fusion. The parallel units in the multi-scale wavelet upsampling layer include a depthwise separable convolutional branch, a double-layer convolutional branch, and a wavelet backpropagation (WPBP) module. The depthwise separable convolutional branch and the double-layer convolutional branch process and enhance the upsampled features through lightweight depthwise separable convolutional layers and standard convolutional layers, ensuring effective aggregation of information across channels and in space. The wavelet backpropagation (WPBP) module is based on the Inverse Discrete Wavelet Transform (IDWT) and introduces a high-frequency information generation mechanism. As the inverse process of the Haar Fast Wavelet Transform, IDWT can adaptively reconstruct high-frequency details discarded in the downsampling stage based on low-resolution features. By combining the retained low-frequency components with the learned high-frequency information, high-quality, near-lossless reconstruction of the feature map is achieved, effectively restoring spatial details and improving the accuracy of subsequent detection and localization.

[0056] For example, such as Figure 5 As shown, the adaptive reparameterization feature enhancement layer includes a multi-branch reparameterization unit, an image enhancement logic branch, and a dynamic feature gating unit. First, the output features of the multi-branch reparameterization unit are fed into the image enhancement logic branch and the dynamic feature gating unit, respectively. Then, the output of the image enhancement logic branch and the output of the multi-branch reparameterization unit are summed by channel dimension, and the result is then multiplied element-wise with the outputs of both the image enhancement logic branch and the dynamic feature gating unit. Through this series of adaptive operations, channel or spatial features highly correlated with small targets can be effectively enhanced, while redundant background information is suppressed, thereby generating output features with stronger discriminative power and a more optimized structure.

[0057] For example, such as Figure 5 As shown, the adaptive reparameterization feature enhancement layer includes a multi-branch reparameterization unit, which comprises three parallel branches. These branches work together to enhance feature representation capabilities through differentiated structural designs. The first branch employs a four-layer parallel convolutional and batch normalization branch structure, consisting of 3×3 convolutions and batch normalization layers, used to extract basic local features of small targets. Subsequently, an asymmetric convolutional structure composed of 3×1 and 1×3 convolutions and batch normalization layers effectively captures edge and texture details in different directions. Finally, a 1×1 convolution and batch normalization layer achieve cross-channel information fusion and dimensional alignment. The second branch consists of depthwise separable convolutions, batch normalization, ReLU activation functions, and 1×1 convolutions, focusing on enhancing the model's non-linear representation capabilities, promoting training convergence, and optimizing channel dimensions. The third branch uses a parallel convolutional group design, including four convolutional structures: 3×3, 3×1, 1×3, and 1×1, further expanding the receptive field of feature extraction. Finally, the outputs of the three branches are fused through element-wise addition to form a comprehensive feature representation that combines multi-scale details and global semantics, significantly improving the model's ability to model small target features in complex scenes.

[0058] For example, such as Figure 5 As shown, the adaptive reparameterization feature enhancement layer also includes an image enhancement logic branch and a dynamic feature gating unit. The image enhancement logic branch introduces an image enhancement logic (IEL) mechanism to adaptively enhance the brightness, contrast, or sharpness of the fused features to compensate for low contrast or blurring in the image, thereby improving the visibility and discriminative power of small targets. Based on Retinex theory, the IEL module decomposes the input brightness features into a brightness basis and reflection, achieving adaptive brightness enhancement through a fusion method of nonlinear activation and element-wise multiplication. The dynamic gating unit generates a set of dynamic weights through 1x1 convolution, batch normalization layers, and a sigmoid activation function. These weights undergo element-wise multiplication to achieve adaptive enhancement and filtering of feature channels or spatial regions.

[0059] For example, the fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer can optionally consist of three cascaded adaptive reparameterization feature enhancement layers. Adjacent adaptive reparameterization feature enhancement layers are dynamically connected to the multi-scale wavelet upsampling and downsampling layers, constructing an adaptive multi-scale feature pyramid. The multi-scale wavelet upsampling layer precisely enhances the spatial resolution of the feature map to meticulously recover and highlight the fine spatial details of small targets. The multi-scale wavelet downsampling layer selectively reduces the feature map resolution, aiming to effectively capture and integrate broader contextual semantic information. This interwoven connection of upsampling and downsampling modules, synergistically with the deep enhancement mechanism of the adaptive reparameterization feature enhancement layer, ensures comprehensive capture and deep fusion of information from the most microscopic local features to the most macroscopic global contextual information, ultimately resulting in deeply fused features.

[0060] Step 160: Input the deep fusion features into the decoder to obtain the small target detection result.

[0061] In object detection models, the decoder is the component responsible for translating abstract features into concrete detection results. Optionally, the decoder is a multi-layer Transformer structure. The deeply fused features are input into this Transformer decoder, and the location bounding boxes of the objects are progressively refined and their class assignments are clarified through a hierarchical processing mechanism. To further optimize the training process, each layer of the decoder is equipped with an auxiliary prediction head to provide intermediate supervision signals. This not only effectively accelerates model convergence but also significantly improves the final detection accuracy through multi-layer feature optimization. The decoder's final output is a series of predicted bounding boxes and their corresponding class confidence scores, completing the transformation from features to detection results.

[0062] This embodiment processes the primary features extracted from the image through convolution and MDPEM modules to enhance the details of small objects, obtaining high-resolution detail features. These high-resolution detail features are then processed through convolution and MDPEM modules to balance detail and semantic information, resulting in medium-resolution balanced features. Finally, these medium-resolution balanced features are processed through a multi-scale wavelet downsampling layer and an MDPEM module to extract deep semantics from small objects and reduce dimensionality, yielding low-resolution semantic features. The MDPEM module comprises a three-layer parallel branch structure consisting of a multi-scale depthwise separable convolutional layer, a depthwise convolutional layer, and an integrated EPGO cue guidance mechanism, used to extract multi-granular spatial context. The system extracts and enhances features at different resolutions, and adjusts feature resolution and information density to achieve multi-scale enhancement and dynamic perception of small target features. The low-resolution semantic features are processed by AIFI to obtain attention-weighted semantic features. These attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features are sequentially input into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features. These deep fusion features are used to integrate semantic information at different resolutions, and the adaptive reparameterization feature enhancement layer further enhances the semantic features of small targets. The deep fusion features are then input into a decoder to obtain the small target detection result. Through this multi-scale feature enhancement and fusion architecture, features at different resolutions, from detail to semantics, are extracted and strengthened, enabling the perception and integration of small target features and ultimately improving the accuracy of small target detection.

[0063] Example 2

[0064] Figure 6 This is a schematic diagram of the structure of the small target detection device based on a neural network provided in Embodiment 2 of the present invention, as shown below. Figure 6 As shown, the device includes:

[0065] The high-resolution detail feature acquisition module 210 is used to process the primary features extracted from the image through convolution and MDPEM modules to enhance the details of small targets and obtain high-resolution detail features. The MDPEM module includes a three-layer parallel branch structure consisting of a multi-scale depth-separable convolutional layer, a depth convolutional layer, and an integrated EPGO cue guidance mechanism. It is used to extract multi-granularity spatial context features and to achieve multi-scale enhancement and dynamic perception of small target features by adjusting the feature resolution and information density.

[0066] The medium-resolution balanced feature acquisition module 220 is used to process the high-resolution detail features through convolution and MDPEM modules to balance details and semantic information, thereby obtaining medium-resolution balanced features. The MDPEM module includes a three-layer parallel branch structure consisting of a multi-scale depth-separable convolutional layer, a depth convolutional layer, and an integrated EPGO prompting mechanism. It is used to extract multi-granularity spatial context features and to achieve multi-scale enhancement and dynamic perception of small target features by adjusting feature resolution and information density.

[0067] The low-resolution semantic feature acquisition module 230 is used to process the medium-resolution balanced features through a multi-scale wavelet downsampling layer and an MDPEM module to extract deep semantics of small targets and reduce their dimensionality, thereby obtaining low-resolution semantic features. The MDPEM module includes a three-layer parallel branch structure consisting of a multi-scale depth-separable convolutional layer, a depth convolutional layer, and an integrated EPGO cue guidance mechanism. It is used to extract multi-granularity spatial context features and to achieve multi-scale enhancement and dynamic perception of small target features by adjusting the feature resolution and information density.

[0068] Attention weighting module 240 is used to process the low-resolution semantic features through AIFI to obtain attention-weighted semantic features;

[0069] The feature fusion module 250 is used to input the attention-weighted semantic features, medium-resolution balanced features and high-resolution detail features into the fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer and an adaptive reparameterization feature enhancement layer in sequence to obtain deep fusion features, which are used to integrate semantic information at different resolutions and enhance the semantic features of small targets using the adaptive reparameterization feature enhancement layer.

[0070] The target detection module 260 is used to input the deep fusion features into the decoder to obtain small target detection results.

[0071] The small target detection device based on neural networks provided in this embodiment enhances small target details by processing primary features extracted from images through convolution and MDPEM modules to obtain high-resolution detail features. These high-resolution detail features are then processed through convolution and MDPEM modules to balance detail and semantic information, resulting in medium-resolution balanced features. Finally, these medium-resolution balanced features are processed through a multi-scale wavelet downsampling layer and an MDPEM module to extract deep semantics of the small targets and reduce their dimensionality, yielding low-resolution semantic features. The MDPEM module includes a three-layer parallel branch structure comprising a multi-scale depth-separable convolutional layer, a depth convolutional layer, and an integrated EPGO cue guidance mechanism, used to improve... Multi-granular spatial context features are extracted, and by adjusting feature resolution and information density, multi-scale enhancement and dynamic perception of small target features are achieved. The low-resolution semantic features are processed using AIFI to obtain attention-weighted semantic features. These attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features are sequentially input into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features. These features integrate semantic information at different resolutions, and the adaptive reparameterization feature enhancement layer further enhances the semantic features of small targets. The deep fusion features are then input into a decoder to obtain the small target detection result. Through this multi-scale feature enhancement and fusion architecture, features at different resolutions, from detail to semantics, are extracted and strengthened, enabling the perception and integration of small target features and ultimately improving the accuracy of small target detection.

[0072] Example 3

[0073] Figure 7 This is a schematic diagram of the structure of a server provided in Embodiment 3 of the present invention. Figure 7 A block diagram is shown of an exemplary server 12 suitable for implementing embodiments of the present invention. Figure 7 The server 12 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0074] like Figure 7 As shown, server 12 is presented in the form of a general-purpose computing server. The components of server 12 may include, but are not limited to: one or more processors or processing units 16, system memory 28, and bus 18 connecting different system components (including system memory 28 and processing unit 16).

[0075] Bus 18 represents one or more of several bus architectures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any of the various bus architectures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect (PCI) bus.

[0076] Server 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by server 12, including volatile and non-volatile media, removable and non-removable media.

[0077] System memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache 32. Server 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, storage system 34 may be used to read and write non-removable, non-volatile magnetic media (… Figure 7 Not shown; usually referred to as a "hard drive"). Although Figure 7 Not shown, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk") and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to bus 18 via one or more data media interfaces. System memory 28 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0078] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28. Such program modules 42 include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. Program modules 42 typically perform the functions and / or methods described in the embodiments of the present invention.

[0079] Server 12 can also communicate with one or more external devices 14 (e.g., keyboard, pointing server, display 24, etc.), and with one or more servers that enable users to interact with server 12, and / or with any server (e.g., network card, modem, etc.) that enables server 12 to communicate with one or more other computing servers. This communication can be performed via input / output (I / O) interface 22. Furthermore, server 12 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 20. As shown, network adapter 20 communicates with other modules of server 12 via bus 18. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with server 12, including but not limited to: microcode, server drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0080] The processing unit 16 executes various functional applications and data processing by running programs stored in the system memory 28, such as implementing the neural network-based small target detection method provided in the embodiments of the present invention.

[0081] Example 4

[0082] Embodiment 4 of the present invention also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform any of the neural network-based small target detection methods provided in the above embodiments.

[0083] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0084] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0085] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0086] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0087] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A small target detection method based on neural networks, characterized in that, include: The primary features extracted from the image are processed by convolution and MDPEM modules to enhance the details of small targets and obtain high-resolution detail features. The high-resolution detail features are processed by convolution and MDPEM modules to balance detail and semantic information, resulting in medium-resolution balanced features. The medium-resolution balanced features are processed by a multi-scale wavelet downsampling layer and an MDPEM module to extract deep semantics of small targets and reduce dimensionality, resulting in low-resolution semantic features. The MDPEM module includes a three-layer parallel branch structure consisting of a multi-scale depthwise separable convolutional layer, a deep convolutional layer, and an integrated EPGO cue guidance mechanism. It is used to extract multi-granularity spatial context features and achieve multi-scale enhancement and dynamic perception of small target features by adjusting feature resolution and information density. The low-resolution semantic features are processed by AIFI to obtain attention-weighted semantic features; The attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features are sequentially input into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features, which are used to integrate semantic information at different resolutions, and the adaptive reparameterization feature enhancement layer is used to enhance the semantic features of small targets. The deep fusion features are input into the decoder to obtain the small target detection results.

2. The method according to claim 1, characterized in that, The multi-scale depth-separable convolutional layer includes: Different receptive field depths can be used to separate convolutional parallel branches to extract multi-granular spatial context information of small targets; The convolutional branch is used to integrate and reduce the dimensionality of features by channel dimension.

3. The method according to claim 2, characterized in that, The deep convolutional layer is used for: The output features of the multi-scale depth-separable convolutional layer are subjected to a first depth-separable convolutional process to obtain a first intermediate feature. The first intermediate feature is subjected to average pooling and a second depthwise separable convolution to obtain the second intermediate feature. The output features, first intermediate features, and second intermediate features of the multi-scale depth separable convolutional layer are summed by channel dimension to obtain features adjusted for spatial resolution and information density.

4. The method according to claim 3, characterized in that, The three-layer parallel branch structure integrating the EPGO prompting and guidance mechanism is used for: The features after adjusting the spatial resolution and information density are input into the first branch, the second branch, and the third branch, respectively. Element-wise dot product is performed on the output features of the second branch and the third branch to obtain the dot product features. The dot product features are summed with the output features of the first branch and the features adjusted for spatial resolution and information density by adding them together by feature dimension to obtain the summed features. The summed features are then convolved to obtain multi-scale enhanced features. The first branch includes the EPGO module, which is used to extract global context information of the features after spatial resolution and information density adjustment, and generate a scaling factor adapted to the channel dimension; The second branch is used for: The features with adjusted spatial resolution and information density are transformed into an information flow through a linear transformation layer, and then a mask between 0 and 1 is generated by ReLU and activation function to adaptively emphasize important feature regions. The third branch is used for: The features with adjusted spatial resolution and information density are passed through a linear transformation layer to form an information flow, and then processed by separable convolution to extract local spatial information of small targets. Finally, an activation function is used to enhance the ability to extract important details of small targets.

5. The method according to claim 1, characterized in that, The multi-scale wavelet downsampling layer includes: Partial convolutional units, pointwise convolutional units, and wavelet forward propagation modules; The partial convolutional unit includes N parallel partial convolutional branches, which are used to extract spatial detail information from multiple local perspectives; The pointwise convolutional unit includes a single pointwise convolutional branch, which is used to integrate small target information focusing on the channel dimension; The wavelet forward propagation module includes a second-order decomposition of wavelet transform, which is used to extract the low-frequency structural information of small targets, avoid high-frequency noise interference, and finally discard the high-frequency subband and retain the low-frequency subband.

6. The method according to claim 1, characterized in that, The multi-scale wavelet upsampling layer includes: Linear mapping with repeated extended paths, linear mapping with channel displacement paths, and parallel elements; The linear mapping and repeated expansion path are used to perform dimensional adaptation and spatial replication of small target features by fusing linear mapping and the repeat function. The linear mapping and channel displacement path, by fusing the linear mapping and channel displacement mechanisms, are used to achieve feature interaction and recombination across channels and space; The parallel unit includes a depthwise separable convolution branch, a two-layer convolution branch, and an integrated inverse discrete wavelet transform branch, which are used to preserve the detailed information of small targets and realize the restoration of the original resolution feature map.

7. The method according to claim 1, characterized in that, The adaptive reparameterization feature enhancement layer includes: Multi-branch reparameterization unit, image enhancement logic branch, and dynamic feature gating unit; The multi-branch reparameterization unit includes four parallel convolution and batch normalization branches, a depthwise separable convolution serial branch, and four parallel convolution branches. The four parallel convolution and batch normalization branches are used to obtain multi-dimensional feature information of small targets, the depthwise separable convolution serial branch is used to enhance the nonlinear expression capability of the model, and the four parallel convolution branches are used to expand the receptive field range of small target feature extraction. The image enhancement logic branch integrates an image enhancement logic module, which is used to enhance the brightness, contrast, and sharpness of features. The dynamic feature gating unit generates a set of weights through convolution, batch normalization layers and activation functions, and multiplies the weights element-wise to achieve adaptive enhancement and filtering of feature channels or spatial regions.

8. A small target detection device based on a neural network, characterized in that, include: The high-resolution detail feature acquisition module is used to process the primary features extracted from the image through convolution and MDPEM modules to enhance the details of small targets and obtain high-resolution detail features. The medium-resolution balanced feature acquisition module is used to process the high-resolution detail features through convolution and MDPEM modules to balance details and semantic information, thereby obtaining medium-resolution balanced features. The low-resolution semantic feature acquisition module is used to process the medium-resolution balanced features through a multi-scale wavelet downsampling layer and an MDPEM module to extract the deep semantics of small targets and reduce their dimensionality, thereby obtaining low-resolution semantic features. The attention-weighted module is used to process the low-resolution semantic features through AIFI to obtain attention-weighted semantic features; The feature fusion module is used to sequentially input the attention-weighted semantic features, medium-resolution balanced features, and high-resolution detail features into a fusion architecture consisting of a multi-scale wavelet upsampling layer, a multi-scale wavelet downsampling layer, and an adaptive reparameterization feature enhancement layer to obtain deep fusion features, which are used to integrate semantic information at different resolutions and enhance the semantic features of small targets using the adaptive reparameterization feature enhancement layer. The target detection module is used to input the deep fusion features into the decoder to obtain small target detection results.

9. A server, characterized in that, The server includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the neural network-based small target detection method as described in any one of claims 1-7.

10. A storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the neural network-based small target detection method as described in any one of claims 1-7.

Images