Wavelet-based full-process enhancement encoder-based target detection method and system

Abstract

The application discloses a target detection method and system based on a wavelet full-process enhanced encoder. The application comprises: after pre-processing a to-be-detected image, inputting the to-be-detected image into a target detection neural network, and outputting a category label and a boundary box coordinate of the target; the target detection neural network comprises a backbone network, an encoder, a decoder and a prediction head connected in sequence; the encoder comprises a wavelet enhancement module and a cross-stage local module; the wavelet enhancement module performs dual-domain enhancement in a wavelet domain and a spatial domain on high-level semantic features output by the backbone network, obtains enhanced semantic features, and performs cross-scale fusion on the enhanced semantic features and other scale features output by the backbone network; and the output of each splicing layer is compensated for boundary information, directional texture and contour information by the cross-stage local module. The application decouples a main structure and directional details through wavelet transformation, effectively suppresses complex background interference, compensates for details prone to degradation in cross-scale propagation, and significantly improves dense target detection precision.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and specifically to a target detection method and system based on a wavelet full-process enhanced encoder. Background Technology

[0002] Object detection is one of the core tasks in computer vision, aiming to locate and identify predefined objects of interest from images or videos. Object detection technology has significant application value in civilian fields such as remote sensing image analysis, intelligent transportation, industrial automation, and natural resource management. In recent years, with the rapid development of high-resolution optical imaging equipment and fixed or mobile ground observation platforms, the demand for technologies that can automatically and accurately perform object detection in complex environments has become increasingly urgent.

[0003] Currently, mainstream general-purpose object detection methods are based on convolutional neural networks or Transformer architectures. These methods perform well on standard datasets such as PASCAL VOC and MS COCO, but they still face significant technical bottlenecks when directly applied to complex environments with densely textured backgrounds and closely arranged objects.

[0004] One issue is the interference from complex backgrounds. When an image contains natural landforms (such as shorelines and bare soil), dense artificial facilities (such as storage yards and building complexes) and dynamic textures (such as water ripples and vegetation) that are similar to the target structure, existing methods mainly enhance high-level semantic features by using attention mechanisms for feature recalibration in the spatial domain. This lacks explicit decoupling of low-frequency main structures from high-frequency directional details. Therefore, it is difficult to accurately suppress complex background interference that is confused with the target's main structure in the frequency domain. This can easily lead to the target's main structural information being overwhelmed by similar background responses, resulting in the target being missed or the background being misjudged as the target.

[0005] Secondly, there is the problem of dense target distribution. For example, in specific environments such as work areas, parking areas, and passageways, multiple targets are arranged in a highly dense distribution, such as being adjacent, end-to-end, or oriented in an alternating manner. The intervals between adjacent targets are extremely narrow, and the targets themselves may have slender or irregularly shaped outlines, with highly similar visual textures. During the cross-scale feature propagation and fusion process of deep networks, these narrow boundaries, fine directional textures, and slender outlines—key discriminative information used to distinguish different individuals—are prone to aliasing or gradual degradation. This makes it difficult for the model to clearly define the independent range of each target, thereby reducing the detection accuracy and recall rate.

[0006] Therefore, how to effectively suppress complex background interference composed of dense textures and similar structures, and maintain the boundary and detail information of the target during cross-scale feature propagation, so as to improve the detection accuracy of densely arranged targets, is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention proposes a target detection method and system based on a wavelet full-process enhanced encoder. The aim is to solve the problems of existing methods in suppressing interference from complex backgrounds, preserving the boundaries of densely distributed targets, and compensating for cross-scale details, thereby effectively improving the accuracy, robustness, and engineering applicability of target detection.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention proposes a target detection method based on a wavelet full-process enhanced encoder, comprising the following steps:

[0010] The target image to be detected is acquired, and the target image is preprocessed to obtain an image tensor;

[0011] The image tensor is input into a pre-trained target detection neural network, which performs target detection on the target image and outputs the category labels and bounding box coordinates of all detected targets.

[0012] The target detection neural network includes a backbone network, an encoder, a decoder, and a prediction head connected in sequence.

[0013] The backbone network is used to extract multi-scale features from the image tensor;

[0014] The encoder includes a wavelet enhancement module and a cross-stage local module;

[0015] The wavelet enhancement module is used to receive high-level semantic features output by the backbone network and enhance the high-level semantic features to obtain enhanced semantic features; the process of enhancing the high-level semantic features by the wavelet enhancement module includes wavelet domain enhancement and collaborative enhancement of global context and local context.

[0016] The enhanced semantic features are fused with other scale features output by the backbone network in a top-down and bottom-up cross-scale manner to obtain multi-scale compensated features; during the cross-scale fusion process, the output of each splicing layer is compensated for boundary information, orientation texture and contour information by a cross-stage local module.

[0017] The multi-scale compensation features are then spliced ​​along the channel dimension to obtain the encoder output;

[0018] The decoder receives the output of the encoder and performs decoding processing to output a target representation vector;

[0019] The prediction head receives the target representation vector and generates the category label and bounding box coordinates based on the target representation vector.

[0020] Furthermore, the wavelet enhancement module includes a wavelet attention branch and a wavelet collaborative enhancement branch;

[0021] The wavelet attention branch is used to decompose the input features into low-frequency subbands and high-frequency subbands using two-dimensional discrete wavelet transform, explicitly decouple the main structure in the low-frequency subband from the directional details in the high-frequency subband, and enhance the high-level semantic representation through subband-level channel and spatial recalibration to obtain wavelet-enhanced features;

[0022] The wavelet collaborative enhancement branch is used to jointly combine the scene-level semantics and local contextual information of the input features in the spatial domain to construct collaboratively enhanced features;

[0023] The enhanced semantic features are obtained by fusing the wavelet enhancement features and the collaborative enhancement features.

[0024] Furthermore, the wavelet attention branch's processing of the input features includes:

[0025] The input features are decomposed into one low-frequency sub-band and three high-frequency sub-bands using two-dimensional discrete wavelet transform; channel enhancement and spatial enhancement are then performed on the four sub-bands respectively.

[0026] The process of channel enhancement for the four sub-bands is as follows: The four sub-bands are concatenated along the channel dimension to obtain a concatenated sub-band. Global average pooling and global max pooling are then applied to the concatenated sub-band to obtain a frequency band-level average statistical description and a frequency band-level maximum statistical description. These two descriptions are then input into a shared multilayer perceptron (MLP) to obtain average pooling features and max pooling features. The sum of these two features is processed by an activation function to generate frequency domain channel weights. These frequency domain channel weights are then divided into four sub-band weights, with each weight applied to its corresponding sub-band, resulting in four channel-enhanced sub-bands.

[0027] The process of spatial enhancement for the four subbands is as follows: each subband is averaged along the channel dimension to obtain the corresponding single-channel spatial response map. The four single-channel spatial response maps are concatenated along the channel dimension and then processed by convolution to generate a shared spatial attention map. The shared spatial attention map is activated by the sigmoid function and then applied to the four subbands to obtain four spatially enhanced subbands.

[0028] The corresponding channel enhancement subband and spatial enhancement subband are fused to obtain four fused subbands; after the four fused subbands are spliced ​​together, they are integrated across subbands to obtain wavelet enhancement features.

[0029] Furthermore, the wavelet collaborative enhancement branch's processing of input features includes:

[0030] Perform global context enhancement and local context enhancement on the input features respectively;

[0031] The process of global context enhancement for input features is as follows: global average pooling is performed on the input features to obtain global pooled features; the global pooled features are compressed using convolution to obtain first compressed features; the first compressed features are sequentially processed by batch normalization, ReLU activation, and convolution to obtain activated compressed features; the activated compressed features are expanded to construct a scene-level semantic representation.

[0032] The process of local context enhancement of input features is as follows: local features are extracted from the input features using convolution to obtain local convolutional features; batch normalization and ReLU activation are performed on the local convolutional features in sequence to obtain local context responses;

[0033] The scene-level semantic representation and the local context response are concatenated in the channel dimension to obtain the concatenated global-local features. The concatenated global-local features are activated by the sigmoid function to obtain activated global-local features. The activated global-local features are multiplied element-wise by the local context response to obtain gated weighted local features. The gated weighted local features are added to the scene-level semantic representation to obtain gated fusion features.

[0034] The input features are weighted to obtain adaptive coefficients. These adaptive coefficients are then multiplied by the gated fusion features to obtain the collaboratively enhanced features.

[0035] Furthermore, the cross-stage local module includes a basic feature extraction branch and a direction-aware wavelet branch;

[0036] The basic feature extraction branch is used to sequentially perform convolution transformation and residual connection processing on the input features to obtain basic features;

[0037] The direction-aware wavelet branch is used to compensate for boundary information, directional texture, and contour information of the input features to obtain directional wavelet enhanced features; the compensation process of the direction-aware wavelet branch for the input features includes: channel compression of the input features through convolution to obtain a second compressed feature;

[0038] Perform a two-dimensional discrete wavelet transform on the second compressed feature to decompose the second compressed feature into one low-frequency sub-band and three high-frequency sub-bands;

[0039] The low-frequency subband is mapped by convolution to obtain a low-frequency representation;

[0040] Different convolution strategies are used to model the three high-frequency subbands in a differentiated manner, with at least two high-frequency subbands employing asymmetric convolution that matches their own directional characteristics, resulting in three directional enhanced high-frequency subbands.

[0041] The three enhanced high-frequency subbands are spliced ​​along the channel dimension to obtain the spliced ​​high-frequency features; the spliced ​​high-frequency features are then mapped and compressed through convolution to obtain the high-frequency representation.

[0042] The high-frequency representation is added to the low-frequency representation to obtain the fused feature;

[0043] The fused features are sequentially upsampled and output projected to obtain directional wavelet enhanced features;

[0044] The basic features are added together with the directional wavelet enhancement features.

[0045] Furthermore, the preprocessing includes size normalization and standardization.

[0046] Furthermore, the decoder is a Transformer decoder.

[0047] This invention also proposes a target detection system based on a wavelet full-process enhanced encoder to implement the above-mentioned target detection method, comprising:

[0048] The image acquisition module is used to acquire the target image to be detected;

[0049] The preprocessing module is used to preprocess the target image to obtain an image tensor;

[0050] The object detection module includes a pre-trained object detection neural network, which is used to perform object detection on the image tensor and output the category labels and bounding box coordinates of all detected objects.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] (1) This invention sets up a wavelet attention branch and a wavelet collaborative enhancement branch in the wavelet enhancement module. The wavelet attention branch uses two-dimensional discrete wavelet transform to explicitly decompose high-level semantic features into one low-frequency sub-band and three high-frequency sub-bands, and enhances the main structure and directional details respectively through sub-band level channel and spatial recalibration. The wavelet collaborative enhancement branch combines scene-level semantics and local context information in the spatial domain to construct collaboratively enhanced features. The output of the wavelet attention branch and the output of the wavelet collaborative enhancement branch are fused, so that the low-frequency main structure and high-frequency directional details are enhanced differentially. Since the wavelet transform decouples the main structure and directional details into different sub-bands, the high-frequency components in the background interference, such as densely textured landforms and regularly arranged artificial facilities, are concentrated in the high-frequency sub-band for processing, while the main structure of the target is preserved in the low-frequency sub-band, and the two are effectively separated in the frequency domain. Therefore, this invention can suppress complex background interference that is confused with the target main structure in the frequency domain, and significantly improve the purity and robustness of the target main representation.

[0053] (2) This invention introduces a direction-aware wavelet branch in the cross-stage local module. This branch performs channel compression on the input features and then performs discrete wavelet transform to obtain one low-frequency sub-band and three high-frequency sub-bands. The main structure of the low-frequency sub-band is stably preserved through convolutional mapping; the three high-frequency sub-bands are modeled differently using a convolutional strategy adapted to the directional characteristics of the sub-bands. This direction-adaptive convolutional strategy can more effectively extract edge and texture features that match the directional characteristics of each sub-band, enhancing the direction-sensitive response. The three enhanced high-frequency sub-bands are concatenated and compressed into a high-frequency representation through a unified mapping, then added to the low-frequency representation and upsampled to restore the resolution, resulting in directional wavelet enhancement features. The directional wavelet enhancement features are injected into the cross-stage local module as compensation signals and added element-wise to the output of the basic feature extraction branch. Since the orientation-aware wavelet branch is specifically designed to compensate for high-frequency details that are prone to degradation during cross-scale propagation, and the direction of compensation matches the inherent orientation characteristics of each sub-band, it can continuously maintain the discrimination information of narrow boundaries, fine orientation textures and slender contours in dense target scenes, significantly enhancing the distinguishability and positioning accuracy between adjacent targets.

[0054] (3) The wavelet enhancement module of this invention only operates on the high-level semantic feature stage of the encoder and does not process all multi-scale features. Therefore, the computational overhead of wavelet transform and subband enhancement is controlled in the key stage of the encoder, avoiding the computational burden brought by the full-process wavelet processing. The direction-aware wavelet branch adopts a lightweight convolution strategy, the channel compression uses 1×1 convolution, and the differential modeling of high-frequency subbands uses 1×7 and 7×1 asymmetric convolution. The number of parameters and the amount of computation for these convolution operations are all less than those of the standard 3×3 convolution. Attached Figure Description

[0055] Figure 1This is a schematic diagram of the overall structure of the target detection neural network in an embodiment of the present invention;

[0056] Figure 2 This is a schematic diagram of the wavelet enhancement module in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the direction-aware wavelet branch in an embodiment of the present invention. Detailed Implementation

[0058] 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.

[0059] Example

[0060] This embodiment proposes a target detection method based on a wavelet full-process enhanced encoder, including the following steps:

[0061] Step 1: Obtain the target image to be detected and preprocess the obtained target image.

[0062] The target image can be a picture taken by a surveillance camera or a high-resolution optical remote sensing image.

[0063] The preprocessing of the target image includes size normalization and standardization. In this embodiment, the size of the input image is uniformly scaled to 640×640 pixels by size normalization. On this basis, standardization is performed to map the pixel value range from [0,255] to [0,1], so that the distribution of image data meets the input requirements of the target detection neural network, thereby obtaining the image tensor.

[0064] Step 2: Input the image tensor into the pre-trained target detection neural network, perform target detection on the target image through the target detection neural network, and output the category labels and bounding box coordinates of all detected targets.

[0065] refer to Figure 1 The target detection neural network is based on the Transformer architecture and includes a backbone network, an encoder, a decoder and a prediction head connected in sequence.

[0066] The backbone network is used to extract multi-scale features from the input image tensor. In this embodiment, the backbone network adopts the backbone feature extractor in the RT-DETR (Real-Time Detection Transformer) baseline model. The extracted multi-scale features include shallow features, mid-level features, and high-level semantic features. Shallow features have high resolution and contain rich edge and texture details, while high-level semantic features have lower resolution but strong semantic information. The resolution and semantic information of mid-level features are between those of shallow and high-level features. In this embodiment, the backbone network extracts three features of different sizes from the input image tensor, denoted as S3, S4, and S5, where S3 is the shallow feature, S4 is the mid-level feature, and S5 is the high-level semantic feature.

[0067] The encoder includes a wavelet enhancement module and a cross-stage local module. The wavelet enhancement module receives high-level semantic features output from the backbone network and enhances these features to obtain enhanced semantic features. These enhanced semantic features are then fused with other scale feature maps output from the backbone network in a top-down and bottom-up cross-scale manner to obtain multi-scale compensated features. These multi-scale compensated features are then concatenated along the channel dimension to obtain the encoder's output.

[0068] Top-down cross-scale fusion is achieved through the FPN (Feature Pyramid Network) structure, while bottom-up cross-scale fusion is achieved through the PAN (Path Aggregation Network) structure. During the cross-scale fusion process, the output of each stitching layer is compensated for boundary information, orientation texture, and contour information by cross-stage local modules.

[0069] In this embodiment, S5 is processed by the wavelet enhancement module to obtain enhanced semantic features. , Perform top-down and bottom-up cross-scale fusion with S4 and S3: After upsampling, it is concatenated with S4 to obtain concatenated feature C45. After upsampling C45, it is concatenated with S3 to obtain concatenated feature C345. C45 is then processed by cross-stage local modules to obtain compensation feature. After C345 is processed by a cross-stage local module, the compensated feature Z3 is obtained; after downsampling Z3, it is compared with... The concatenation is performed to obtain concatenated feature C34. After downsampling C34, it is combined with the enhanced semantic features. The splicing process yields spliced ​​features C5 and C34, which are then processed by cross-stage local modules to obtain compensation features. After C5 undergoes cross-stage local module processing, the compensation features Z5 and Z3 are obtained. After splicing Z5 along the channel dimension, the fusion compensation feature is obtained, which is the output of the encoder.

[0070] It should be noted that, in order to unify the channel dimensions for subsequent processing, 1×1 convolutions can be used to adjust the channels of the multi-scale feature maps output by the backbone network before inputting them into the encoder, depending on the actual needs.

[0071] The high-level semantic features output by the backbone network are used as input features for the wavelet enhancement module, referencing... Figure 2 The wavelet enhancement module includes a wavelet attention branch and a wavelet collaborative enhancement branch.

[0072] The enhancement process of high-level semantic features by the wavelet attention branch is as follows:

[0073] The input features are decomposed into four sub-bands using two-dimensional discrete wavelet transform: the first LL sub-band, the first LH sub-band, the first HL sub-band, and the first HH sub-band. Channel enhancement and spatial enhancement are then performed on each of the four sub-bands.

[0074] The channel enhancement process for the four sub-bands is as follows: The four sub-bands are spliced ​​along the channel dimension to obtain spliced ​​sub-bands. Global average pooling and global max pooling are then applied to the spliced ​​sub-bands to obtain a frequency band-level average statistical description and a frequency band-level maximum statistical description. These descriptions are then input into a shared multilayer perceptron to obtain average pooling features and max pooling features. The average pooling features and max pooling features are summed element-wise and activated by a sigmoid function to generate frequency domain channel weights. These weights are then divided into four sub-band weights based on the number of channels in each original sub-band. Each sub-band weight is applied to its corresponding original sub-band, resulting in four channel-enhanced sub-bands: channel-enhanced LL sub-band, channel-enhanced LH sub-band, channel-enhanced HL sub-band, and channel-enhanced HH sub-band. This channel enhancement process enables adaptive recalibration of the sub-bands in the wavelet domain, thereby enhancing the representation ability of the target structure and effectively suppressing complex background interference that may confuse the target structure.

[0075] The spatial enhancement process for the four subbands is as follows: For each subband, average aggregation is performed along the channel dimension, i.e., the spatial mean of all channels in each subband is taken to obtain the corresponding single-channel spatial response map. The four single-channel spatial response maps are then concatenated along the channel dimension to obtain a concatenated spatial response map. This concatenated spatial response map is then processed using a 7×7 convolution to obtain a shared spatial attention map. This shared spatial attention map is activated by the sigmoid function and applied to the four original subbands, resulting in four spatially enhanced subbands: Spatial Enhancement LL subband, Spatial Enhancement LH subband, Spatial Enhancement HL subband, and Spatial Enhancement HH subband. The spatial enhancement process preserves the spatial location information within the subbands, further enhancing the directional details and edge texture of the target.

[0076] The corresponding channel enhancement subband and spatial enhancement subband are added element-wise to obtain four fused subbands. The four fused subbands are then spliced ​​along the channel dimension and then sequentially processed by 1×1 convolution, batch normalization, ReLU activation and upsampling to integrate across subbands, resulting in wavelet enhancement features.

[0077] The enhancement process of the wavelet collaborative enhancement branch for high-level semantic features includes: performing global context enhancement and local context enhancement on the high-level semantic features respectively;

[0078] The process of global context enhancement for high-level semantic features is as follows: Global average pooling is performed on the high-level semantic features to obtain global pooled features; the global pooled features are compressed using 1×1 convolutions, reducing the number of channels to 1 / r of the original number, resulting in the first compressed feature, where r is the compression ratio (r=4 in this embodiment); the first compressed feature is then subjected to batch normalization, ReLU activation, and 1×1 convolution processing sequentially to obtain the activated compressed feature; the activated compressed feature is then expanded to restore the original number of channels, constructing a scene-level semantic representation. This global context enhancement process can extract scene-level semantic information from the entire image, enabling the model to understand the overall background of the image, thereby distinguishing between the target and the background at a high-level semantic level.

[0079] The process of local context enhancement for high-level semantic features is as follows: local features are extracted from high-level semantic features using 3×3 convolution to obtain local convolutional features; batch normalization and ReLU activation are performed on the local convolutional features in sequence to obtain local context response.

[0080] The scene-level semantic representation and the local context response are concatenated along the channel dimension to obtain the concatenated global-local features. The concatenated global-local features are activated by the sigmoid function to obtain activated global-local features. The activated global-local features are multiplied element-wise by the local context response to obtain gated weighted local features. The gated weighted local features are added element-wise by the scene-level semantic representation to obtain gated fusion features.

[0081] The high-level semantic features are weighted to obtain adaptive coefficients. The weight generation process is as follows: the high-level semantic features are sequentially subjected to global average pooling to obtain channel statistical features; the channel statistical features are then sequentially subjected to linear mapping dimensionality reduction, ReLU activation, linear mapping dimensionality increase, and sigmoid activation to generate adaptive coefficients.

[0082] The adaptive coefficients are multiplied by the gated fusion features to obtain the collaboratively enhanced features. The collaborative enhancement process dynamically fuses global semantics and local details through a learnable gating mechanism, achieving adaptive feature enhancement and further improving the distinguishability of adjacent targets in dense target scenes.

[0083] The wavelet enhancement features output by the wavelet attention branch and the collaborative enhancement features output by the wavelet collaborative enhancement branch are added element-wise to obtain the output of the wavelet enhancement module, which is the enhanced semantic feature.

[0084] The cross-stage local module includes a basic feature extraction branch and a direction-aware wavelet branch. In this embodiment, the basic feature extraction branch adopts a CSP (Cross Stage Partial Network) structure.

[0085] The basic feature extraction branch is used to sequentially perform convolution transformation and residual connection processing on the input features to obtain basic features;

[0086] The direction-aware wavelet branch is used to compensate the input features for boundary information, direction texture and contour information to obtain directional wavelet enhanced features;

[0087] The basic features are added element-wise to the directional wavelet enhancement features to obtain the compensation features.

[0088] refer to Figure 3 The compensation process for input features by the direction-aware wavelet branch includes:

[0089] The input features are compressed by 1×1 convolution, reducing the number of channels to half of the original number, resulting in the second compressed feature. This channel compression operation reduces the computational cost of subsequent wavelet transform and convolution processing.

[0090] Perform a two-dimensional discrete wavelet transform on the second compressed feature to decompose the compressed feature into four sub-bands: the second LL sub-band, the second LH sub-band, the second HL sub-band, and the second HH sub-band;

[0091] The low-frequency representation is obtained by mapping the second LL subband using 1×1 convolution. This operation stably preserves the main structural information of the target.

[0092] Different convolution strategies are used to perform differentiated modeling for the second LH subband, the second HL subband, and the second HH subband. At least two subbands use asymmetric convolution that matches their own directional characteristics. In this embodiment, the second LH subband uses 1×7 asymmetric convolution, the second HL subband uses 7×1 asymmetric convolution, and the second HH subband uses 3×3 convolution, resulting in directionally enhanced LH subband, directionally enhanced HL subband, and directionally enhanced HH subband. Through this directionally adapted convolution strategy, edge and texture features that match the directional characteristics of the subbands can be effectively extracted, enhancing the directionally sensitive response.

[0093] The directional enhancement LH subband, directional enhancement HL subband, and directional enhancement HH subband are spliced ​​along the channel dimension to obtain the spliced ​​high-frequency features; the spliced ​​high-frequency features are then uniformly mapped and compressed using 1×1 convolution to obtain a high-frequency representation; this operation fuses the high-frequency information from the three directions into a compact high-frequency feature representation.

[0094] The high-frequency representation and the low-frequency representation are added element by element to obtain the fused features; this fusion operation organically combines the main structural information with the directional detail information.

[0095] The fused features are sequentially upsampled and projected to restore spatial resolution and adjust channel dimensions, resulting in directional wavelet-enhanced features. In this embodiment, bilinear interpolation is used for upsampling, and 3×3 convolution is used for output projection. The directional wavelet-enhanced features serve as compensation signals to enhance boundary, directional texture, and contour information in the stitching layer output.

[0096] The decoder receives the fused compensation features output by the encoder, performs target query and feature interaction through multi-head self-attention mechanism and cross-attention mechanism, gradually optimizes the target representation, and outputs the target representation vector; in this embodiment, the decoder adopts the Transformer decoder structure.

[0097] After processing the target representation vector output by the decoder, the prediction head outputs the category labels and bounding box coordinates of all detected targets.

[0098] In this embodiment, the training and validation of the object detection neural network can utilize public datasets such as ShipRSImageNet or Seaships7000. Fine-grained object detection tasks are trained and validated on the ShipRSImageNet dataset, while the generalization ability of object detection is trained and validated on the Seaships7000 dataset. The appropriate dataset can be selected based on specific needs in practical applications. During training, the Adam optimizer is used, with cross-entropy loss function for classification loss and L1 loss function or generalized intersection-union loss function for regression loss. The network parameters are iteratively updated using the backpropagation algorithm until the model converges.

[0099] To verify the effectiveness of the present invention, the present invention uses RT-DETR as the baseline model and constructs three improved models based on RT-DETR: improved RT-DETR-1, improved RT-DETR-2 and improved RT-DETR-3.

[0100] The difference between the improved RT-DETR-1 and RT-DETR is that the encoder in RT-DETR is replaced with the encoder in the above embodiment.

[0101] The difference between the improved RT-DETR-2 and the improved RT-DETR-1 is that the encoder of the improved RT-DETR-2 does not contain a wavelet enhancement module.

[0102] The difference between the improved RT-DETR-3 and the improved RT-DETR-1 is that the encoder of the improved RT-DETR-3 does not contain cross-stage local modules.

[0103] RT-DETR, improved RT-DETR-1, improved RT-DETR-2, and improved RT-DETR-3 were validated on the ShipRSImageNet and Seaships7000 datasets, respectively, with an evaluation metric of mAP50:95. The ShipRSImageNet dataset contains four levels: level 0 to level 3. This invention validates the performance of each model at level 2 (subclass granularity) and level 3 (specific model granularity). On the same dataset, the four models were validated under the same input size and training strategy. The validation results are shown in Table 1.

[0104] Table 1 Performance Comparison of Each Model

[0105]

[0106] As shown in Table 1, the improved RT-DETR-1 achieves the best mAP50:95 score across all three evaluation metrics, with the most significant improvement observed in the fine-grained classification task. This indicates that the present invention, by explicitly decoupling the low-frequency main structure from the high-frequency directional details through a wavelet enhancement module and by compensating for easily degraded boundary and texture information during cross-scale propagation through direction-aware wavelet branching, can more effectively improve the model's recognition capability in complex backgrounds and densely distributed scenes.

[0107] The specific embodiments of the present invention are provided to enable those skilled in the art to understand or implement 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.

[0108] It should be understood that the present invention is not limited to the content already described above, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims.

Claims

1. A target detection method based on a wavelet full-process enhancement encoder, characterized in that, Includes the following steps: The target image to be detected is acquired, and the target image is preprocessed to obtain an image tensor; The image tensor is input into a pre-trained target detection neural network, which performs target detection on the target image and outputs the category labels and bounding box coordinates of all detected targets. The target detection neural network includes a backbone network, an encoder, a decoder, and a prediction head connected in sequence. The backbone network is used to extract multi-scale features from the image tensor; The encoder includes a wavelet enhancement module and a cross-stage local module; The wavelet enhancement module is used to receive high-level semantic features output by the backbone network and enhance the high-level semantic features to obtain enhanced semantic features; the process of enhancing the high-level semantic features by the wavelet enhancement module includes wavelet domain enhancement and collaborative enhancement of global context and local context. The enhanced semantic features are fused with other scale features output by the backbone network in a top-down and bottom-up cross-scale manner to obtain multi-scale compensated features; during the cross-scale fusion process, the output of each splicing layer is compensated for boundary information, orientation texture and contour information by a cross-stage local module. The multi-scale compensation features are spliced ​​along the channel dimension to obtain the encoder output; The decoder receives the output of the encoder and performs decoding processing to output a target representation vector; The prediction head receives the target representation vector and generates the category label and bounding box coordinates based on the target representation vector.

2. The target detection method based on the wavelet full-process enhancement encoder according to claim 1, characterized in that, The wavelet enhancement module includes a wavelet attention branch and a wavelet collaborative enhancement branch; The wavelet attention branch is used to decompose the input features into low-frequency subbands and high-frequency subbands using two-dimensional discrete wavelet transform, explicitly decouple the main structure in the low-frequency subband from the directional details in the high-frequency subband, and enhance the high-level semantic representation through subband-level channel and spatial recalibration to obtain wavelet-enhanced features; The wavelet collaborative enhancement branch is used to jointly combine the scene-level semantics and local contextual information of the input features in the spatial domain to construct collaboratively enhanced features; The enhanced semantic features are obtained by fusing the wavelet enhancement features and the collaborative enhancement features.

3. The target detection method based on the wavelet full-process enhancement encoder according to claim 2, characterized in that, The wavelet attention branch's processing of input features includes: The input features are decomposed into one low-frequency sub-band and three high-frequency sub-bands using two-dimensional discrete wavelet transform; channel enhancement and spatial enhancement are then performed on the four sub-bands respectively. The process of channel enhancement for the four sub-bands is as follows: the four sub-bands are spliced ​​along the channel dimension to obtain spliced ​​sub-bands. Global average pooling and global max pooling are then performed on the spliced ​​sub-bands to obtain a frequency band-level average statistical description and a frequency band-level maximum statistical description. These two descriptions are then input into a shared multilayer perceptron to obtain average pooling features and max pooling features. The sum of these features is processed by an activation function to generate frequency domain channel weights. These frequency domain channel weights are then divided into four sub-band weights, with each weight applied to its corresponding sub-band to obtain four channel-enhanced sub-bands. The process of spatial enhancement for the four subbands is as follows: each subband is averaged along the channel dimension to obtain the corresponding single-channel spatial response map. The four single-channel spatial response maps are concatenated along the channel dimension and then processed by convolution to generate a shared spatial attention map. The shared spatial attention map is activated by the sigmoid function and then applied to the four subbands to obtain four spatially enhanced subbands. The corresponding channel enhancement subband and spatial enhancement subband are fused to obtain four fused subbands; after the four fused subbands are spliced ​​together, they are integrated across subbands to obtain wavelet enhancement features.

4. The target detection method based on a wavelet full-process enhanced encoder according to claim 2, characterized in that, The wavelet collaborative enhancement branch's processing of input features includes: Perform global context enhancement and local context enhancement on the input features respectively; The process of global context enhancement for input features is as follows: global average pooling is performed on the input features to obtain global pooled features; the global pooled features are compressed using convolution to obtain first compressed features; the first compressed features are sequentially processed by batch normalization, ReLU activation, and convolution to obtain activated compressed features; the activated compressed features are expanded to construct a scene-level semantic representation. The process of local context enhancement of input features is as follows: local features are extracted from the input features using convolution to obtain local convolutional features; batch normalization and ReLU activation are performed on the local convolutional features in sequence to obtain local context responses; The scene-level semantic representation and the local context response are concatenated in the channel dimension to obtain the concatenated global-local features. The concatenated global-local features are activated by the sigmoid function to obtain activated global-local features. The activated global-local features are multiplied element-wise by the local context response to obtain gated weighted local features. The gated weighted local features are added to the scene-level semantic representation to obtain gated fusion features. The input features are weighted to obtain adaptive coefficients. These adaptive coefficients are then multiplied by the gated fusion features to obtain the collaboratively enhanced features.

5. The target detection method based on a wavelet full-process enhanced encoder according to claim 1, characterized in that, The cross-stage local module includes a basic feature extraction branch and a direction-aware wavelet branch; The basic feature extraction branch is used to sequentially perform convolution transformation and residual connection processing on the input features to obtain basic features; The direction-aware wavelet branch is used to compensate the input features for boundary information, direction texture and contour information to obtain directional wavelet enhanced features; The compensation process for input features by the direction-aware wavelet branch includes: channel compression of the input features through convolution to obtain a second compressed feature; Perform a two-dimensional discrete wavelet transform on the second compressed feature to decompose the second compressed feature into one low-frequency sub-band and three high-frequency sub-bands; The low-frequency sub-band is mapped by convolution to obtain a low-frequency representation; Different convolution strategies are used to model the three high-frequency subbands in a differentiated manner, with at least two high-frequency subbands employing asymmetric convolution that matches their own directional characteristics, resulting in three directional enhanced high-frequency subbands. The three enhanced high-frequency subbands are spliced ​​along the channel dimension to obtain the spliced ​​high-frequency features; the spliced ​​high-frequency features are then mapped and compressed through convolution to obtain the high-frequency representation. The high-frequency representation is added to the low-frequency representation to obtain the fused feature; The fused features are sequentially upsampled and output projected to obtain directional wavelet enhanced features; The basic features are added together with the directional wavelet enhancement features.

6. The target detection method based on a wavelet full-process enhanced encoder according to claim 1, characterized in that, The preprocessing includes size normalization and standardization.

7. The target detection method based on a wavelet full-process enhanced encoder according to claim 1, characterized in that, The decoder is a Transformer decoder.

8. A target detection system based on a wavelet full-process enhanced encoder, used to implement the target detection method according to any one of claims 1 to 7, characterized in that, include: The image acquisition module is used to acquire the target image to be detected; The preprocessing module is used to preprocess the target image to obtain an image tensor; The object detection module includes a pre-trained object detection neural network, which is used to perform object detection on the image tensor and output the category labels and bounding box coordinates of all detected objects.

Images