A small target detection method and system based on high-frequency information guidance

Abstract

The application discloses a kind of small target detection method and system based on high frequency information guidance, it is related to computer vision and deep learning technical field.The method comprises the following steps: collecting the original image of target area, image enhancement is obtained based on frequency domain guidance undersampling, image multistage backbone feature is obtained by using backbone network feature extraction, alignment fusion feature is obtained based on alignment fusion strategy before fusion, based on fine-grained interactive learning strategy, fine-grained difference attention map is constructed using competitive weight allocation mechanism and image alignment fusion feature, residual enhancement is carried out on the original image using fine-grained difference attention map, and finally enhanced detection feature is obtained and input into detection prediction branch, and small target detection result is output.The application realizes the comprehensive improvement of detection precision, robustness and universality of regional small target detection through the synergistic effect of frequency domain guidance undersampling, alignment before fusion and fine-grained interactive learning.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and deep learning technology, and in particular to a method and system for small target detection based on high-frequency information guidance. Background Technology

[0002] Current mainstream object detection methods (such as the YOLO series and RT-DETR) mainly focus on utilizing spatial domain features, identifying objects by learning low-frequency semantic information such as shape and contour. Common improvement strategies include feature fusion (such as FPN), optimizing network architecture, and introducing attention mechanisms. However, these methods do not fully utilize the high-frequency information contained in the frequency domain and still have the following shortcomings: 1. Insufficient utilization of high-frequency information: The discriminative features of small targets (such as edges and textures) are mainly concentrated in high-frequency components, but existing methods often ignore high-frequency information in the frequency domain.

[0003] 2. Severe feature loss: During multiple downsampling processes, deep neural networks will irreversibly lose high-frequency components above the Nyquist frequency according to the sampling theorem, resulting in a lack of sufficient discriminative power in deep features.

[0004] 3. Feature alignment ambiguity: Traditional nearest neighbor upsampling is prone to semantic ambiguity and inaccurate spatial localization when restoring resolution.

[0005] Therefore, proposing a small target detection method to accurately capture high-frequency information in the frequency domain, alleviate the irreversible loss of high-frequency components in deep neural network downsampling, and solve the semantic ambiguity and inaccurate localization problems of traditional upsampling are technical problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a small target detection method and system based on high-frequency information guidance. By utilizing high-frequency information in the frequency domain to guide small target detection, it solves the technical defects of existing small target detection such as insufficient utilization of high-frequency information, serious downsampling loss, inaccurate feature fusion, and weak fine-grained recognition, and comprehensively improves the detection accuracy, robustness and versatility of small target detection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A small target detection method based on high-frequency information guidance includes the following steps: S1 Acquisition Steps: Acquire the original image of the target area; S2 enhancement steps: The original image is enhanced based on frequency domain guided downsampling to obtain the enhanced image; S3 Extraction Steps: Use a backbone network to extract features from the enhanced image to obtain multi-level backbone features of the image; S4 fusion step: Align and fuse multi-level backbone features of the image based on the pre-fusion alignment strategy to obtain image alignment and fusion features; S5 construction steps: Based on a fine-grained interactive learning strategy, a fine-grained differential attention map is constructed using a competitive weight allocation mechanism and image alignment fusion features; S6 Residual Step: The original image is augmented using a fine-grained differential attention map to obtain the final enhanced detection features; S7 Detection Steps: Input the final enhanced detection features into the detection prediction branch and output the small target detection results.

[0008] Optionally, S2 of the above method includes the following steps: S21 Separation Step: The high-frequency components of the original image are obtained by separating the high-frequency and low-frequency components of the original image through discrete wavelet transform; S22 Filtering Step: The adaptive dual-gating mechanism is used to preserve the high-frequency details of the high-frequency components of the original image and filter image noise to obtain the enhanced image.

[0009] Optionally, S4 of the above method includes the following steps: S41 Transform Enhancement Steps: After applying wavelet transform to the shallow skeletal features of the multi-level skeletal features of the image, the shallow high-frequency components are extracted, and the directional details are enhanced using strip attention to obtain the enhanced shallow skeletal features. S42 Alignment and Fusion Steps: After fusing the enhanced shallow backbone features and the deep semantic features of the multi-level backbone features of the image, inverse wavelet transform is performed for alignment to obtain the image alignment and fusion features.

[0010] Optionally, the above method, S5, includes the following steps: S51 normalization step: The image alignment and fusion features are normalized using the Softmax function to generate a difference attention map; S52 Construction and Allocation Steps: Construct a competitive weight allocation mechanism and perform weight allocation processing on the differential attention map to obtain a fine-grained differential attention map.

[0011] Optionally, in the above method, the small target detection results in the S7 detection step include the small target's category, location confidence, and detection box.

[0012] A small target detection system based on high-frequency information guidance includes a data acquisition module, an FGD module, an extraction module, an FBAL module, an FGIL module, a residual enhancement module, and a detection output module connected in sequence. The acquisition module is used to acquire raw images of the target area; The FGD module is used to perform image enhancement on the original image based on frequency domain guided downsampling to obtain the enhanced image; The extraction module is used to extract features from the enhanced image using the backbone network to obtain multi-level backbone features of the image; The FBAL module is used to align and fuse multi-level backbone features of an image based on a pre-fusion alignment strategy to obtain image alignment and fusion features. The FGIL module is used to construct fine-grained differential attention maps based on a fine-grained interactive learning strategy, utilizing a competitive weight allocation mechanism and image alignment fusion features. The residual enhancement module is used to perform residual enhancement on the original image using fine-grained difference attention maps to obtain the final enhanced detection features; The detection output module is used to input the final enhanced detection features into the detection prediction branch and output the small target detection results.

[0013] In the aforementioned system, optionally, the FGD module includes interconnected separation submodules and filtering submodules; The separation submodule is used to separate the high-frequency and low-frequency components of the original image through discrete wavelet transform to obtain the high-frequency component of the original image. The filtering submodule is used to preserve high-frequency details of the high-frequency components of the original image and filter image noise using an adaptive dual-gating mechanism to obtain an enhanced image.

[0014] In the aforementioned system, optionally, the FBAL module includes interconnected transform enhancement submodules and alignment fusion submodules; The transform enhancement submodule is used to apply wavelet transform to the shallow skeletal features of the multi-level skeletal features of the image, extract the shallow high-frequency components, and use strip attention to enhance directional details to obtain the enhanced shallow skeletal features. The alignment and fusion submodule is used to fuse the enhanced shallow backbone features and the deep semantic features of the multi-level backbone features of the image, and then perform inverse wavelet transform alignment to obtain the image alignment and fusion features.

[0015] The above system, optionally, includes an FGIL module comprising interconnected normalization submodules and build allocation submodules; The normalization step is used to normalize the image alignment and fusion features using the Softmax function, generating a difference attention map; The construction and allocation step is used to build a competitive weight allocation mechanism and perform weight allocation processing on the differential attention map to obtain a fine-grained differential attention map.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention provides a small target detection method and system based on high-frequency information guidance, which has the following beneficial effects: (1) Significantly improved accuracy: By introducing a high-frequency guidance mechanism, high-frequency discrimination information is introduced into small target detection, which greatly improves the small target discrimination ability; on challenging small target datasets such as VisDrone and UAVDT, it has achieved a significant improvement in detection performance compared with the benchmark model; (2) Reduce information loss: Compared with traditional downsampling, the present invention uses frequency domain guided downsampling to minimize the loss of high-frequency information from the source, providing a more robust feature input for subsequent detection; (3) More accurate positioning: The present invention adopts a fusion pre-alignment strategy, which introduces high-frequency details to guide feature reconstruction, solves the semantic ambiguity problem in deep networks, and makes the target boundary positioning more accurate; (4) Reduced false negative rate: The differential attention mechanism of the fine-grained interactive learning strategy significantly enhances fine-grained learning, and the false negative rate and false positive rate of small targets are significantly reduced. (5) Integrability: The framework based on the content disclosed in this invention has strong versatility and can be flexibly integrated into various existing detectors based on FPN architecture (such as YOLOv11, D-FINE, etc.), and the computational overhead is controllable, which significantly reduces the detection cost. Attached Figure Description

[0017] 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a small target detection method guided by high-frequency information disclosed in this invention. Figure 2 This is a schematic diagram illustrating the operating principle of frequency-domain guided downsampling as disclosed in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the operational principle of the pre-fusion alignment disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the operational principle of fine-grained interactive learning as disclosed in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0021] See Figure 1 As shown, this invention discloses a small target detection method based on high-frequency information guidance, comprising the following steps: S1 Acquisition Steps: Acquire the original image of the target area; S2 enhancement steps: The original image is enhanced based on frequency domain guided downsampling to obtain the enhanced image; S3 Extraction Steps: Use a backbone network to extract features from the enhanced image to obtain multi-level backbone features of the image; S4 fusion step: Align and fuse multi-level backbone features of the image based on the pre-fusion alignment strategy to obtain image alignment and fusion features; S5 construction steps: Based on a fine-grained interactive learning strategy, a fine-grained differential attention map is constructed using a competitive weight allocation mechanism and image alignment fusion features; S6 Residual Step: The original image is augmented using a fine-grained differential attention map to obtain the final enhanced detection features; S7 Detection Steps: Input the final enhanced detection features into the detection prediction branch and output the small target detection results.

[0022] like Figure 1 As shown, this invention is a high-frequency guided small target detection framework based on the Feature Pyramid Network (FPN) architecture. It achieves accurate detection of small targets through frequency domain guided downsampling, fusion pre-alignment, and fine-grained interactive learning. It can be directly embedded into detectors of various Feature Pyramid Network (FPN) architectures. The implementation of accurate small target detection will be further explained below from three aspects: frequency domain guided downsampling, fusion pre-alignment, and fine-grained interactive learning.

[0023] Optionally, S2 specifically includes the following steps: S21 Separation Step: The high-frequency components of the original image are obtained by separating the high-frequency and low-frequency components of the original image through discrete wavelet transform; S22 Filtering Step: The adaptive dual-gating mechanism is used to preserve the high-frequency details of the high-frequency components of the original image and filter image noise to obtain the enhanced image.

[0024] Unlike existing methods, this invention employs Frequency-Guided Downsampling (FGD) to process the original image during the initial feature extraction stage. In the S2 enhancement step, reference... Figure 2 This invention, based on Functional Gradient Diffraction (FGD), separates high- and low-frequency components using discrete wavelet transform and employs an adaptive dual-gating mechanism to preserve high-frequency details while filtering noise. Compared to existing feature extraction processes, FGD enables this invention to obtain enhanced backbone features with complete preservation of low- and high-frequency information in the S2 extraction step.

[0025] Optionally, S4 specifically includes the following steps: S41 Transform Enhancement Steps: After applying wavelet transform to the shallow skeletal features of the multi-level skeletal features of the image, the shallow high-frequency components are extracted, and the directional details are enhanced using strip attention to obtain the enhanced shallow skeletal features. S42 Alignment and Fusion Steps: After fusing the enhanced shallow backbone features and the deep semantic features of the multi-level backbone features of the image, inverse wavelet transform is performed for alignment to obtain the image alignment and fusion features.

[0026] Combination Figure 1 and Figure 3 After the input image is enhanced by frequency domain guided downsampling, the features are extracted by the backbone network to obtain multi-level backbone features (F2, F3, F4, F5).

[0027] Based on the multi-level backbone features of the image, when constructing the enhanced feature pyramid, this invention further designs a Fuse Before Alignment (FBAL) process to replace traditional upsampling. The FBAL processing designed in the S4 fusion step of this invention applies wavelet transform to the shallow backbone features of F3, F4, and F5 to extract horizontal / vertical / diagonal high-frequency components, enhances directional details through bar attention, and then fuses with deep semantic features through inverse wavelet transform alignment, ultimately obtaining fused features (P5, P4, P3) with precise alignment of spatial and semantic features. During the fusion process, F2 has no effect; it is the preliminary result of feature extraction from the enhanced input image by the backbone network. Subsequent multi-level backbone features and fused features are all extracted or fused based on F2.

[0028] Optionally, S5 specifically includes the following steps: S51 normalization step: The image alignment and fusion features are normalized using the Softmax function to generate a difference attention map; S52 Construction and Allocation Steps: Construct a competitive weight allocation mechanism to perform weight allocation processing on the difference attention map, and obtain a fine-grained difference attention map.

[0029] Specifically, after obtaining the aligned fusion features in step S4, the precisely aligned fusion features are processed using the Fine-Grained Interactive Learning (FGIL) strategy in step S5. (See reference...) Figure 4 First, a Softmax function is applied to the fused features to generate a differential attention map. Then, a competitive weight allocation mechanism is constructed to obtain a fine-grained differential attention map. Due to the competitive weight relationship inherent in the fine-grained differential attention map, this invention can analyze the image from two different perspectives, thereby effectively improving the invention's ability to detect small targets.

[0030] Optionally, the small target detection results in the S7 detection step include the small target's category, location confidence, and detection box.

[0031] Fine-grained differential attention maps are used to perform residual enhancement on the original fused features, which strengthens the fine-grained feature learning of weak response regions of small targets, resulting in the final enhanced detection features. The final enhanced detection features are then input into the detection prediction branch, which outputs the category, location confidence, and detection box of the small target, ultimately achieving accurate detection of small targets.

[0032] The present invention also discloses methods for implementing such Figure 1 The method shown is a small target detection system based on high-frequency information guidance, which includes a data acquisition module, an FGD module, an extraction module, an FBAL module, an FGIL module, a residual enhancement module, and a detection output module connected in sequence.

[0033] The acquisition module is used to acquire raw images of the target area; The FGD module is used to perform image enhancement on the original image based on frequency domain guided downsampling to obtain the enhanced image; The extraction module is used to extract features from the enhanced image using the backbone network to obtain multi-level backbone features of the image; The FBAL module is used to align and fuse multi-level backbone features of an image based on a pre-fusion alignment strategy to obtain image alignment and fusion features. The FGIL module is used to construct fine-grained differential attention maps based on a fine-grained interactive learning strategy, utilizing a competitive weight allocation mechanism and image alignment fusion features. The residual enhancement module is used to perform residual enhancement on the original image using fine-grained difference attention maps to obtain the final enhanced detection features; The detection output module is used to input the final enhanced detection features into the detection prediction branch and output the small target detection results.

[0034] Combination Figure 1 The input image is processed by the backbone network's feature extraction module. During the feature extraction stage, the FGD module is used to reduce high-frequency loss. When constructing the enhanced feature pyramid, the FBAL module replaces the traditional upsampling and uses shallow high-frequency features to guide the reconstruction of deep semantic features. The FGIL module performs fine-grained enhancement on the fused features and finally outputs the detection results.

[0035] Optionally, the FGD module includes interconnected separation and filtering sub-modules; The separation submodule is used to separate the high-frequency and low-frequency components of the original image through discrete wavelet transform to obtain the high-frequency component of the original image. The filtering submodule is used to preserve high-frequency details of the high-frequency components of the original image and filter image noise using an adaptive dual-gating mechanism to obtain an enhanced image.

[0036] Optionally, the FBAL module includes interconnected transform enhancement submodules and alignment blending submodules; The transform enhancement submodule is used to apply wavelet transform to the shallow skeletal features of the multi-level skeletal features of the image, extract the shallow high-frequency components, and use strip attention to enhance directional details to obtain the enhanced shallow skeletal features. The alignment and fusion submodule is used to fuse the enhanced shallow backbone features and the deep semantic features of the multi-level backbone features of the image, and then perform inverse wavelet transform alignment to obtain the image alignment and fusion features.

[0037] Optionally, the FGIL module includes interconnected normalization submodules and build allocation submodules; The normalization step is used to normalize the image alignment and fusion features using the Softmax function, generating a difference attention map; The construction and allocation step is used to build a competitive weight allocation mechanism to perform weight allocation processing on the differential attention map, resulting in a fine-grained differential attention map.

[0038] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A small target detection method based on high-frequency information guidance, characterized in that, Includes the following steps: S1 Acquisition Steps: Acquire the original image of the target area; S2 enhancement steps: The original image is enhanced based on frequency domain guided downsampling to obtain the enhanced image; S3 Extraction Steps: Use a backbone network to extract features from the enhanced image to obtain multi-level backbone features of the image; S4 fusion step: Align and fuse multi-level backbone features of the image based on the pre-fusion alignment strategy to obtain image alignment and fusion features; S5 construction steps: Based on a fine-grained interactive learning strategy, a fine-grained differential attention map is constructed using a competitive weight allocation mechanism and image alignment fusion features; S6 Residual Step: The original image is augmented using a fine-grained differential attention map to obtain the final enhanced detection features; S7 Detection Steps: Input the final enhanced detection features into the detection prediction branch and output the small target detection results.

2. The small target detection method based on high-frequency information guidance according to claim 1, characterized in that, S2 specifically includes the following steps: S21 Separation Step: The high-frequency components of the original image are obtained by separating the high-frequency and low-frequency components of the original image through discrete wavelet transform; S22 Filtering Step: The adaptive dual-gating mechanism is used to preserve the high-frequency details of the high-frequency components of the original image and filter image noise to obtain the enhanced image.

3. The small target detection method based on high-frequency information guidance according to claim 1, characterized in that, S4 specifically includes the following steps: S41 Transform Enhancement Steps: After applying wavelet transform to the shallow skeletal features of the multi-level skeletal features of the image, the shallow high-frequency components are extracted, and the directional details are enhanced using strip attention to obtain the enhanced shallow skeletal features. S42 Alignment and Fusion Steps: After fusing the enhanced shallow backbone features and the deep semantic features of the multi-level backbone features of the image, inverse wavelet transform is performed for alignment to obtain the image alignment and fusion features.

4. The small target detection method based on high-frequency information guidance according to claim 1, characterized in that, S5 specifically includes the following steps: S51 normalization step: The image alignment and fusion features are normalized using the Softmax function to generate a difference attention map; S52 Construction and Allocation Steps: Construct a competitive weight allocation mechanism and perform weight allocation processing on the differential attention map to obtain a fine-grained differential attention map.

5. The small target detection method based on high-frequency information guidance according to claim 1, characterized in that, The small target detection results in the S7 detection step include the small target's category, location confidence, and detection box.

6. A small target detection system based on high-frequency information guidance, characterized in that, It includes a data acquisition module, an FGD module, an extraction module, an FBAL module, an FGIL module, a residual enhancement module, and a detection output module connected in sequence; The acquisition module is used to acquire raw images of the target area; The FGD module is used to perform image enhancement on the original image based on frequency domain guided downsampling to obtain the enhanced image; The extraction module is used to extract features from the enhanced image using the backbone network to obtain multi-level backbone features of the image; The FBAL module is used to align and fuse multi-level backbone features of an image based on a pre-fusion alignment strategy to obtain image alignment and fusion features. The FGIL module is used to construct fine-grained differential attention maps based on a fine-grained interactive learning strategy, utilizing a competitive weight allocation mechanism and image alignment fusion features. The residual enhancement module is used to perform residual enhancement on the original image using fine-grained difference attention maps to obtain the final enhanced detection features; The detection output module is used to input the final enhanced detection features into the detection prediction branch and output the small target detection results.

7. A small target detection system based on high-frequency information guidance according to claim 6, characterized in that, The FGD module includes interconnected separation and filtering sub-modules; The separation submodule is used to separate the high-frequency and low-frequency components of the original image through discrete wavelet transform to obtain the high-frequency component of the original image. The filtering submodule is used to preserve high-frequency details of the high-frequency components of the original image and filter image noise using an adaptive dual-gating mechanism to obtain an enhanced image.

8. A small target detection system based on high-frequency information guidance according to claim 6, characterized in that, The FBAL module consists of interconnected transform enhancement submodules and alignment blending submodules; The transform enhancement submodule is used to apply wavelet transform to the shallow skeletal features of the multi-level skeletal features of the image, extract the shallow high-frequency components, and use strip attention to enhance directional details to obtain the enhanced shallow skeletal features. The alignment and fusion submodule is used to fuse the enhanced shallow backbone features and the deep semantic features of the multi-level backbone features of the image, and then perform inverse wavelet transform alignment to obtain the image alignment and fusion features.

9. A small target detection system based on high-frequency information guidance according to claim 6, characterized in that, The FGIL module includes interconnected normalization submodules and build allocation submodules; The normalization step is used to normalize the image alignment and fusion features using the Softmax function, generating a difference attention map; The construction and allocation step is used to build a competitive weight allocation mechanism to perform weight allocation processing on the differential attention map, resulting in a fine-grained differential attention map.

Images