Image information bidirectional guiding fusion method, target detection method and system

Abstract

This invention discloses a bidirectional guided fusion method for image information, a target detection method, and a system. The fusion method includes: inputting the image information to be processed into a YOLO model to obtain a shallow feature image from the Backbone network and a deep feature image from the Neck network; decoupling the feature frequency bands of the shallow feature image and the upsampled deep feature image to obtain a shallow high-frequency feature map and a deep low-frequency feature map, respectively; analyzing the semantic information represented by the deep low-frequency feature map to guide the shallow high-frequency feature map to enhance details, resulting in a detail-enhanced image; evaluating the weights of each feature channel using the shallow high-frequency feature map to guide the deep low-frequency feature map to calibrate semantics, resulting in a calibrated image; and fusing the detail-enhanced image with the calibrated image. This invention solves the problems of detail loss and semantic ambiguity in existing target detection networks, especially for small-scale targets.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a bidirectional guided fusion method for image information, a target detection method, and a system. Background Technology

[0002] In real-world engineering scenarios such as drone aerial photography, high-altitude security, and traffic inspection, the targets being detected (pedestrians, vehicles, and small facilities) have a very small pixel ratio in the image, weak texture information, and are easily submerged by complex backgrounds, making them typical small-scale targets.

[0003] Existing technologies generally employ feature pyramids such as FPN and PAN for feature fusion, but they all use simple channel splicing or element addition. The fusion process is static and non-adaptive, resulting in the following drawbacks: the fusion method is coarse, failing to distinguish the essential differences between deep semantics and shallow details, leading to low efficiency in information complementarity; background noise in shallow features is amplified synchronously, and effective details of small targets cannot be specifically enhanced; and there is a lack of a two-way guidance mechanism, meaning that deep semantics cannot guide shallow denoising, and shallow details cannot calibrate deep semantics, resulting in high false negative rates, inaccurate positioning, and poor robustness for small targets in scenarios such as aerial photography.

[0004] Single-stage object detectors, represented by the YOLO series, are widely used due to their balance between speed and accuracy. However, they have significant bottlenecks in small object detection. The fundamental reason lies in the hierarchical structure of convolutional neural networks: deep features have strong semantic information but low spatial resolution, resulting in the loss of a large amount of small object details; shallow features have high spatial resolution and preserve details, but weak discriminative semantics and severe background noise interference.

[0005] In summary, in real-world scenarios such as drone aerial photography and high-altitude monitoring, traditional feature pyramids cannot simultaneously preserve details of small targets, suppress noise, and adapt to multiple scales, making it difficult to meet the engineering deployment requirements of real-time, high precision, and high robustness.

[0006] The disclosure of the above background technical content is only for the purpose of assisting in understanding the concept and technical solution of this application, and does not necessarily provide technical instruction. Summary of the Invention

[0007] The purpose of this invention is to provide a bidirectional guided fusion method for image information applicable to small-scale targets, solving the problems of loss of detailed information and semantic ambiguity in existing target detection networks, especially for small-scale targets.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A bidirectional guided fusion method for image information includes the following steps: The image information to be processed is input into the YOLO model to obtain the shallow feature image from the Backbone network of the YOLO model and the deep feature image from the Neck network with the PAN structure. The shallow feature image is decoupled by feature frequency band to obtain the corresponding shallow low-frequency feature map C. L-low and shallow high-frequency feature map C L-high ; The deep feature image is upsampled to align with the shallow feature image, and the feature frequency bands of the upsampled deep feature image are decoupled to obtain the corresponding deep low-frequency feature map P. D-low and deep high-frequency feature map P D-high ; Analysis of the deep low-frequency feature map P D-low The semantic information represented is used to guide the shallow high-frequency feature map C. L-high Enhance details to obtain a detail-enhanced image C L-high-Enhance ; and utilize the shallow high-frequency feature map C L-high The weights of each feature channel are evaluated to guide the deep low-frequency feature map P. D-low The calibration semantics are used to obtain the calibration image P. D-low-Calibrate ; The detailed enhanced image C L-high-Enhance With calibration image P D-low-Calibrate Integration.

[0009] Furthermore, following any one or a combination of the aforementioned technical solutions, the detail enhancement image C L-high-Enhance With calibration image P D-low-Calibrate Scale-adaptive fusion includes: The detailed enhanced image C L-high-Enhance and calibration image P D-low-Calibrate Create temporary channels for stitching to obtain a stitched image; Analyze the scene-dependent features of the stitched image to determine high-frequency and low-frequency dependencies; Based on high-frequency dependence and low-frequency dependence, the detail-enhanced image C is... L-high-Enhance and calibration image P D-low-Calibrate Assign weights to each: if the high frequency dependency is higher, then the detail-enhanced image C is given a higher weight. L-high-Enhance The larger the assigned weight value, the higher the low-frequency dependence, and the better the calibration image P. D-low-Calibrate The larger the assigned weight value; The detail-enhanced image C is then processed according to the assigned weight values. L-high-Enhance With calibration image P D-low-Calibrate Weighted fusion.

[0010] Furthermore, following any one or a combination of the aforementioned technical solutions, a pre-trained adaptive weight generation model is used to enhance the detail of the image C. L-high-Enhance Assign weight values α And for calibrating image P D-low-Calibrate Assign weight values β : The adaptive weight generation model applies the stitched image Global average pooling is performed to compress the spatial features of each channel into a channel-level statistical scalar, resulting in the channel feature vector. ;in, H To stitch images X height, W To stitch images X width, C C Increase the number of channels in the image to enhance detail. C P To calibrate the number of channels in the image; The feature vectors at each channel level are input into an attention network consisting of a neural network and an activation function; During training, the attention network updates its parameters through backpropagation of the loss function, configuring it to: identify high-frequency dependency features represented by small objects in the image and low-frequency dependency features represented by large objects in the image; and output weight values ​​for significant high-frequency dependency features. α Tends to be greater than the weight value β For features with significant low-frequency dependence, the output weight value is... β Tends to be greater than the weight value α .

[0011] Furthermore, following any one or a combination of the aforementioned technical solutions, the attention network generates a weight vector corresponding to the channel-level feature vector z through a nonlinear activation function: ,in, s The weight vector corresponding to the channel-level feature vector z. σ And δ are nonlinear activation functions, W 1 and W 2 represents a learnable parameter; For the weight vector s The front of the middle C C Each corresponding detail-enhanced image C L-high-Enhance The average of the weighted elements is used as the weight value. α ; For the weight vector s The middle C POne corresponding calibration image P D-low-Calibrate The average of the weighted elements is used as the weight value. β .

[0012] Furthermore, following any one or a combination of the aforementioned technical solutions, the feature image is decoupled from its feature frequency band in the following manner, wherein the feature image is the shallow feature image or the upsampled deep feature image: The feature image is subjected to low-pass filtering to obtain a low-frequency feature map; and the feature image and the low-frequency feature map are then subjected to a difference operation to obtain a high-frequency feature map. Alternatively, the feature image can be high-pass filtered to obtain a high-frequency feature map; and the feature image and the high-frequency feature map can be differentially analyzed to obtain a low-frequency feature map.

[0013] Furthermore, following any one or a combination of the aforementioned technical solutions, the feature image is decoupled from its feature frequency band in the following manner, wherein the feature image is the shallow feature image or the upsampled deep feature image: The feature image is transformed from the spatial domain to the frequency domain to obtain the spectral information corresponding to the feature image. The spectral information corresponding to the feature image is subjected to low-pass filtering to obtain low-frequency spectral information; and the spectral information corresponding to the feature image and the low-frequency spectral information are differentially processed to obtain high-frequency spectral information; or, the spectral information corresponding to the feature image is subjected to high-pass filtering to obtain high-frequency spectral information; and the spectral information corresponding to the feature image and the high-frequency spectral information are differentially processed to obtain low-frequency spectral information. The low-frequency spectrum information is inversely transformed to obtain a low-frequency feature map in the spatial domain; the high-frequency spectrum information is inversely transformed to obtain a high-frequency feature map in the spatial domain.

[0014] Furthermore, following any one or a combination of the aforementioned technical solutions, the deep low-frequency feature map P is obtained through the following method. D-low The semantic information of the representation guides the shallow high-frequency feature map C L-high Enhanced details: For the deep low-frequency feature map P D-low Semantic confidence analysis was performed to obtain the deep low-frequency feature map P. D-low The semantic confidence of each spatial location; A semantic guidance mask is generated, wherein the element values ​​are mapped to a preset continuous interval by the semantic confidence of each spatial location to represent the semantic saliency of the corresponding spatial location. The semantic guidance mask is used to analyze the shallow high-frequency feature map C. L-high Weighting is performed to obtain the detail-enhanced image C.L-high-Enhance .

[0015] Furthermore, following any one or a combination of the aforementioned technical solutions, semantic confidence analysis is achieved through a semantic analysis network, wherein the semantic analysis network comprises a convolutional layer and an activation function layer connected in sequence; The convolutional layer is used to reduce the number of channels in the deep low-frequency feature map to 1, thus obtaining a single-channel feature map. The activation function layer is configured to use the Sigmoid function to map the feature values ​​of each spatial location in the single-channel feature map to the (0,1) interval, thereby obtaining a semantic confidence map; The method further includes: generating the semantic guidance mask based on the semantic confidence map, including: directly using the semantic confidence map as the semantic guidance mask; or, performing linear scaling or nonlinear transformation on the semantic confidence map to obtain the semantic guidance mask.

[0016] Furthermore, following any one or a combination of the aforementioned technical solutions, the shallow high-frequency feature map C is obtained through the following method. L-high Guided deep low-frequency feature map P D-low Calibration semantics: Based on the shallow high-frequency feature map C L-high Calculate the detail evaluation vector that represents the detail richness of each feature channel; Map the detailed evaluation vector to a channel weight vector; The deep low-frequency feature map P is analyzed using the channel weight vector. D-low Channel weighting is performed to obtain the calibration image P. D-low-Calibrate .

[0017] Furthermore, based on any one or a combination of the aforementioned technical solutions, the detailed evaluation vector is calculated in the following manner: The shallow high-frequency feature map C L-high It is a three-dimensional tensor, whose dimensions include channel dimension, height dimension, and width dimension, and its eigenvalues ​​are defined. x c_m ( i , j ) is a three-dimensional tensor with channel index of m The x-axis is i The vertical axis is j scalar element at; For channel index cm The feature channels are analyzed globally along the height and width, and the variance, absolute value, mean, or energy are used to determine the first feature channel. m Feature statistics of each feature channel U m ; The feature statistics of all feature channels are combined in channel index order to obtain the detailed evaluation vector. U detail : U detail =[ U 1, U 2,…, U m ,…, U C ],in, U 1 represents the feature statistics of the first feature channel. U 2 represents the feature statistics of the second feature channel. C This represents the total number of feature channels. U C For the first C Feature statistics for each feature channel.

[0018] Furthermore, following any one or a combination of the aforementioned technical solutions, the detailed evaluation vector is mapped to a channel weight vector in the following manner: The element values ​​of the detail evaluation vector are normalized to obtain the normalized detail evaluation vector; The normalized detail evaluation vector is input into a weight mapping network, which includes at least a fully connected layer and an activation function layer. The fully connected layer is configured to perform a linear transformation and offset on the normalized detail evaluation vector, and the activation function layer is configured to use a sigmoid function to map the output of the fully connected layer to the (0,1) interval to obtain the channel weight vector.

[0019] Furthermore, in accordance with any or a combination of the aforementioned technical solutions, the bidirectional guided fusion method for image information provided by the present invention further includes: The deep low-frequency feature map P is adjusted through 1×1 convolution mapping. D-low The number of channels is equal to that of the shallow high-frequency feature map C. L-high The number of channels; Deep low-frequency feature map P after adjusting the number of channels D-low The calibration image P is obtained by performing a channel-by-channel dot product with the channel weight vector. D-low-Calibrate .

[0020] Furthermore, following any one or a combination of the aforementioned technical solutions, the shallow feature image is the feature image located in the penultimate layer of the Backbone network, and the deep feature image is the feature image output from the last layer of the PAN path aggregation Neck network.

[0021] According to another aspect of the present invention, a target detection method is provided, comprising the following steps: Acquire image information; The acquired image information is fused and enhanced using the bidirectional guided fusion method described above. Based on the fused and enhanced image, identify, locate, or analyze target objects.

[0022] According to another aspect of the present invention, an image processing system is provided, comprising the following modules: The feature extraction module is configured to receive image information to be processed and input it into the YOLO model to obtain shallow feature images from the Backbone network of the YOLO model and deep feature images from the Neck network with a PAN structure. An upsampling module is configured to upsample the deep feature image to align it with the shallow feature image. The frequency band decoupling module is configured to perform feature frequency band decoupling on the shallow feature image to obtain the corresponding shallow low-frequency feature map C. L-low and shallow high-frequency feature map C L-high ; and perform feature frequency band decoupling on the upsampled deep feature image to obtain the corresponding deep low-frequency feature map P. D-low and deep high-frequency feature map P D-high ; The detail enhancement guidance module is configured to analyze the deep low-frequency feature map P. D-low The semantic information represented is used to guide the shallow high-frequency feature map C. L-high Enhance details to obtain a detail-enhanced image C L-high-Enhance ; The semantic calibration guidance module is configured to utilize the shallow high-frequency feature map C L-high The weights of each feature channel are evaluated to guide the deep low-frequency feature map P. D-low The calibration semantics are used to obtain the calibration image P. D-low-Calibrate ; The fusion module is configured to fusion the detail-enhanced image C L-high-Enhance With calibration image P D-low-Calibrate The features are then fused to output the final fused feature map.

[0023] According to another aspect of the present invention, a target detection system is provided, including an image acquisition device and an image processing system as described above, wherein the image acquisition device is configured to acquire image information containing a target object; The image processing system is configured to fuse and enhance image information containing the target to be detected; and to identify, locate or analyze the target object based on the fused and enhanced image.

[0024] The beneficial effects of the technical solution provided by this invention are as follows: a. This invention proposes a novel frequency domain decoupling and bidirectional guided fusion mechanism, which breaks the limitations of the traditional PAN structure's unidirectional and coarse fusion method of "upsampling + splicing", so as to more accurately improve the detection performance of small targets (dependent on details) and large targets (dependent on semantics) in complex scenes. b. By decoupling the feature frequency band, the high-frequency detail components (CL-high) in the shallow features are separated from the low-frequency semantic components (PD-low) in the deep features. The deep semantic information is used to guide the enhancement of shallow details, which effectively strengthens the high-frequency information such as edges and textures that are crucial to small targets, and greatly reduces the false detection and missed detection rates of small targets. c. Innovatively utilize the richness of shallow high-frequency details to evaluate and calibrate the channel weights of deep semantic features, enabling the network to adaptively focus on semantic channels that correspond to real details, suppress "hollow" or vague semantic responses, thereby improving the expression quality of semantic features and the accuracy of classification and localization of large and medium-scale targets. d. Based on frequency domain decoupling, guidance and fusion can separate and suppress noise in the feature map to a certain extent (usually present in high-frequency components), and use reliable semantic information to enhance effective details, thereby improving the robustness of the entire feature pyramid; e. The improvements are mainly focused on the Neck part of feature fusion. The core innovative modules (such as frequency band decoupling and adaptive weight generation unit) are designed to be lightweight. While significantly improving performance, they do not bring excessive computation and parameter overhead to the model, which is conducive to deployment and application in resource-constrained scenarios such as mobile devices and edge devices. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating a bidirectional guided fusion method for image information provided as an exemplary embodiment of the present invention; Figure 2 A deep low-frequency feature map P is provided as an exemplary embodiment of the present invention. D-low The semantic information of the representation guides the shallow high-frequency feature map C L-high Enhanced flowchart; Figure 3A shallow high-frequency feature map C is provided as an exemplary embodiment of the present invention. L-high Guided deep low-frequency feature map P D-low A flowchart illustrating the calibration semantics; Figure 4 Detail-enhanced image C provided as an exemplary embodiment of the present invention L-high-Enhance With calibration image P D-low-Calibrate A flowchart illustrating the adaptive fusion process; Figure 5 A schematic flowchart of a target detection method provided as an exemplary embodiment of the present invention; Figure 6 A schematic block diagram of the module structure of an image processing system provided as an exemplary embodiment of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0029] In traditional object detection networks, the original PAN directly aligns and simply adds or concatenates the upsampled deep features with the shallow features. This is equivalent to simply stacking the "semantic map" and the "detail map," failing to distinguish the essential differences between deep semantic features and shallow detail features. This fusion method is often not sensitive enough to weak small target signals (high-frequency details), and they are easily overwhelmed by semantic information during fusion. Furthermore, during the fusion process, on the one hand, the background noise of shallow features is amplified simultaneously, while the effective details of small targets are not enhanced in a targeted manner; on the other hand, strong deep semantic features cannot accurately guide the selection and enhancement of shallow features, and high-resolution shallow details cannot effectively supplement the localization information of deep features, resulting in insufficient cross-level feature complementarity.

[0030] This invention no longer directly fuses the entire feature map. Instead, it decouples the feature map into two parts: "contour details (high frequency)" and "content semantics (low frequency)." These two parts then guide each other before being fused. This invention uses a systematic "decoupling-bidirectional guidance-fusion" framework to separate high-frequency detail components in shallow features from low-frequency semantic components in deep features. It utilizes deep semantic information to guide the enhancement of shallow details and leverages the richness of shallow high-frequency details to evaluate and calibrate deep semantic features. This achieves more accurate and efficient utilization of multi-scale features, providing a high-performance, robust solution for target detection tasks, especially for complex scenes containing small targets.

[0031] In one embodiment of the present invention, a bidirectional guided fusion method for image information is provided, such as... Figure 1 As shown, it includes the following steps: S100: Input the image information to be processed into the YOLO model to obtain the shallow feature image from the Backbone network of the YOLO model and the deep feature image from the Neck network with PAN structure; The purpose of this step is to extract multi-scale features from the input image that are suitable for subsequent bidirectional guided fusion. The image information can be a single RGB image, an image sequence, or a video frame. Before input, the image can be preprocessed by standardization, such as normalizing the pixel values ​​to the [0,1] range, and adjusting the image size (e.g., scaling, padding) to the input size preset by the model (e.g., 640×640 pixels).

[0032] The Backbone network is the backbone of the YOLO model. It is typically a deep convolutional neural network, such as CSPDarknet53 or EfficientNet, and includes multiple downsampling stages. The middle convolutional layers of the Backbone network are responsible for extracting and outputting shallow feature images. These shallow feature images have high spatial resolution and rich detail (such as edges, corners, and textures), but lower semantic abstraction. In this embodiment, the shallow feature image is selected from the penultimate layer output of the Backbone network, such as the C4 feature map in the YOLOv5 / v8 architecture.

[0033] The C4 feature map extracted by the Backbone network is fed into the Neck network, which has a path aggregation network structure, i.e., a PAN structure. This structure, through operations including upsampling, downsampling, and lateral connections, defines the high-level or final output of the Neck network as a deep feature image. Deep feature images have lower spatial resolution but contain rich, high-level semantic information (e.g., object categories, parts, and overall concepts) that has undergone multiple layers of abstraction and fusion, while spatial detail information is relatively limited. In this embodiment, the deep feature image is the output of the last layer of the PAN structure, such as the P5 feature map in YOLOv5 / v8.

[0034] S200: Decouple the feature frequency bands of the shallow feature image to obtain the corresponding shallow low-frequency feature map C. L-low and shallow high-frequency feature map C L-high ; Frequency band decoupling is typically achieved through Fast Fourier Transform (FFT) or Gaussian / Laplace pyramids, mapping spatial features to the frequency domain. Low frequencies correspond to smooth regions, i.e., shallow low-frequency feature maps C. L-low Characterizes smooth backgrounds, basic textures, and other features; high frequencies correspond to areas of drastic change, thus separating the shallow high-frequency feature map C. L-high It characterizes features such as edges, noise, and small target contours.

[0035] S300: Upsample the deep feature image to align it with the shallow feature image, and decouple the feature frequency bands of the upsampled deep feature image to obtain the corresponding deep low-frequency feature map P. D-low and deep high-frequency feature map P D-high ; The purpose of this step is to align the spatial dimensions of deep features and shallow features, and to decompose the aligned deep features into semantically dominant low-frequency components and detail-residual high-frequency components, providing a clean semantic source for subsequent bidirectional guidance.

[0036] Since the spatial size (e.g., 20×20) of deep feature images from the Neck network (such as P5) is typically smaller than that of shallow feature images from the Backbone network (such as C4, 40×40), an upsampling operation needs to be performed on the deep feature images to make their height H and width W consistent with those of the shallow feature images for easier subsequent fusion. The upsampling operation can be implemented using interpolation algorithms, including but not limited to nearest neighbor interpolation, bilinear interpolation, or transposed convolution. After upsampling, a size-aligned deep feature map is obtained, denoted as P. D-aligned .

[0037] Then, for the aligned deep feature map P D-aligned Feature frequency band decoupling is performed, separating it into low-frequency components and high-frequency components. The low-frequency components correspond to stable and global semantic information in deep features, while the high-frequency components correspond to residual and local details and noise information.

[0038] Whether performing feature band decoupling on shallow feature images or on aligned deep feature images, any of the following methods can be used: Method 1 is a spatial domain filtering method. For example, a Gaussian kernel with a standard deviation of σ is used for convolution to perform low-pass filtering on the feature image to obtain a low-frequency feature map; then the feature image and the low-frequency feature map are differentially processed to obtain a high-frequency feature map; or the feature image is subjected to spatial domain Gaussian high-pass filtering to obtain a high-frequency feature map; then the feature image and the high-frequency feature map are differentially processed to obtain a low-frequency feature map.

[0039] Two-dimensional frequency domain filtering methods, such as performing a fast Fourier transform on the feature map and then performing low-pass filtering in the frequency domain: performing a frequency domain transformation on the feature image to convert it from the spatial domain to the frequency domain, and obtaining the spectral information corresponding to the feature image; The spectral information corresponding to the feature image is subjected to low-pass filtering to obtain low-frequency spectral information; and the spectral information corresponding to the feature image and the low-frequency spectral information are differentially processed to obtain high-frequency spectral information; or, the spectral information corresponding to the feature image is subjected to high-pass filtering to obtain high-frequency spectral information; and the spectral information corresponding to the feature image and the high-frequency spectral information are differentially processed to obtain low-frequency spectral information. The low-frequency spectrum information is inversely transformed to obtain a low-frequency feature map in the spatial domain, namely, a deep low-frequency feature map P. D-low The high-frequency spectrum information is inversely transformed to obtain a high-frequency feature map in the spatial domain, namely, a deep high-frequency feature map P. D-high .

[0040] Whether using method one or method two, the deep low-frequency feature map P obtained by filtering D-lowIt is the main carrier of deep semantic information, with gentle spatial changes and rich in high-level semantic features such as categories and components in the channel dimension; while the deep high-frequency feature map P D-high It contains residual edge responses, subtle textures, and possible high-frequency noise from deep features, and its information intensity is usually much weaker than the shallow high-frequency feature map C decoupled from shallow features. L-high .

[0041] Thus, through the above steps S100-S300, a deep low-frequency feature map P with perfectly aligned image spatial dimensions has been obtained. D-low and shallow high-frequency feature map C L-high This completed the "purification" of deep and shallow features.

[0042] This invention does not limit the order of steps S200 and S300. The introduction of frequency band decoupling addresses the coarse-grained problem of homogenizing heterogeneous features (deep semantics and shallow details) in conventional methods by actively decomposing the input feature map into low-frequency and high-frequency components. The low-frequency components focus on the overall contour and semantics, while the high-frequency components focus on edges and details. This provides a categorized and purposeful feature base for subsequent intelligent fusion, rather than directly mixing the original features.

[0043] S400: Analyze the deep low-frequency feature map P D-low The semantic information represented is used to guide the shallow high-frequency feature map C. L-high Enhance details to obtain a detail-enhanced image C L-high-Enhance ; and utilize the shallow high-frequency feature map C L-high The weights of each feature channel are evaluated to guide the deep low-frequency feature map P. D-low The calibration semantics are used to obtain the calibration image P. D-low-Calibrate ; This step is the core step of the present invention, and the former guides and embodies the utilization of deep low-frequency feature maps P. D-low Strong semantic information (such as heatmaps) is used to deactivate shallow high-frequency feature maps C. L-high The corresponding regions in the model allow it to focus more on semantically meaningful details and suppress irrelevant background noise. The latter guides the use of shallow high-frequency feature maps C L-high The richness of detail is used to evaluate channel weights for deep low-frequency feature maps P. D-low The semantic channels are recalibrated: those semantic channels containing detailed information are strengthened, while "hollow" semantic channels are weakened, thus solving the problem of "fuzzy localization" of deep features.

[0044] To address the problem of low complementary efficiency caused by the lack of semantic guidance for details and feedback from details to semantics in simple fusion, two guiding paths are established: the first is a semantic-guided detail path, which uses deep low-frequency semantic information to generate a spatial weight map to enhance (target region) or suppress (noise background) shallow high-frequency detail features, achieving denoising and detail enhancement; the second is a detail-calibrated semantic path, which uses the detail richness of shallow high-frequency details to fine-tune the channel weights of deep low-frequency semantic features, making them more focused on feature channels containing effective targets, thus improving the accuracy of semantic expression.

[0045] The following is a detailed description of the guided implementation with enhanced details, such as... Figure 2 As shown, the deep low-frequency feature map P is obtained through the following method. D-low The semantic information of the representation guides the shallow high-frequency feature map C L-high Enhanced details: S411: For the deep low-frequency feature map P D-low Semantic confidence analysis was performed to obtain the deep low-frequency feature map P. D-low The semantic confidence of each spatial location; S412: Generate a semantic guidance mask, wherein the element values ​​map the semantic confidence of each spatial location to a preset continuous interval to characterize the semantic saliency of the corresponding spatial location. Specifically, a mask is generated using a 1×1 convolution and a sigmoid activation function. Regions with high response values ​​in the deep feature layers represent semantically salient regions. This mask is not a binarized mask. In this embodiment, semantic confidence analysis is implemented through a semantic analysis network, which includes sequentially connected convolutional layers and activation function layers. The convolutional layers reduce the number of channels in the deep low-frequency feature map to 1, obtaining a single-channel feature map. The activation function layers are configured to use the sigmoid function to map the feature values ​​of each spatial location in the single-channel feature map to the (0,1) interval, obtaining a semantic confidence map. The closer the value in the semantic confidence map is to 1, the higher the confidence that the target semantic feature exists at that spatial location; the closer the value is to 0, the higher the confidence that the location is background or has no semantic information. Generating the semantic guidance mask based on the semantic confidence map includes: directly using the semantic confidence map as the semantic guidance mask, i.e., also using the continuous interval (0,1); or, performing linear scaling or nonlinear transformation on the semantic confidence map to obtain the semantic guidance mask M. s .

[0046] S413: Utilize the semantic guidance mask to process the shallow high-frequency feature map C L-high Weighting is performed to obtain the detail-enhanced image C. L-high-EnhanceSpecifically, using pixel-by-pixel multiplication: C L-high-Enhance =M s ⊙C L-high .

[0047] The following is a detailed explanation of the guided implementation of semantic calibration. The shallow high-frequency feature map is responsible for capturing low-level details such as edges and textures, and usually has a large number of channels (e.g., 256 dimensions). When the number of channels in the shallow high-frequency feature map is inconsistent with that in the deep low-frequency feature map, a 1×1 convolution is used to adjust the number of channels in the deep feature map to be consistent with that in the shallow layer. First, configure the convolution kernel parameters: configure a 1×1 convolutional layer with a kernel weight matrix W' of dimension (C). out C in ,1,1), where C in The deep low-frequency feature map P D-low The number of input channels, C out The preset target number of channels (and the shallow high-frequency feature map C) L-high (The number of channels is consistent with 256). Secondly, the deep low-frequency feature map P D-low The input is fed into this 1×1 convolutional layer, for each spatial location ( i , j Linear combination of the eigenvectors of ) P D-low '[:, i , j ]= W'·P D-low [:, i , j ]+b, where b is an optional bias vector.

[0048] This gives us a channel dimension of C. out Adjusted deep low-frequency feature map P D-low '(The number of its channels is the same as that of the shallow high-frequency feature map C)' L-high Deep low-frequency feature map P after adjusting the number of channels D-low The calibration image P is obtained by performing a channel-by-channel dot product with the channel weight vector below. D-low-Calibrate .

[0049] like Figure 3 As shown, the shallow high-frequency feature map C is obtained through the following method. L-high Guided deep low-frequency feature map P D-low Calibration semantics: S421: Based on the shallow high-frequency feature map C L-high Calculate the detail evaluation vector representing the detail richness of each feature channel. U detail Specifically, the shallow high-frequency feature map C L-highIt is a three-dimensional tensor, whose dimensions include channel dimension, height dimension, and width dimension, and its eigenvalues ​​are defined. x c_m ( i , j ) is a three-dimensional tensor with channel index of m The x-axis is i The vertical axis is j scalar element at; For channel index cm The feature channels are analyzed globally along the height and width, and the variance, absolute value, mean, or energy are used to determine the first feature channel. m Feature statistics of each feature channel U m : The variance statistics are as follows: Calculate the first... m Mean of each feature channel Where H is the shallow high-frequency feature map C L-high The height of W represents the shallow high-frequency feature map C. L-high The width; then calculate the first... m The variance of each feature channel is used as a feature statistic. U m : .

[0050] The absolute value mean is statistically analyzed as follows: .

[0051] Energy statistics are as follows: .

[0052] The feature statistics of all feature channels are combined in channel index order to obtain the detailed evaluation vector. U detail : U detail =[ U 1, U 2,…, U m ,…, U C ],in, U 1 represents the feature statistics of the first feature channel. U 2 represents the feature statistics of the second feature channel. C This represents the total number of feature channels. U C For the first C Feature statistics for each feature channel.

[0053] S422: Map the detail evaluation vector to a channel weight vector; specifically, normalize the values ​​of each element of the detail evaluation vector to obtain a normalized detail evaluation vector; input the normalized detail evaluation vector into a weight mapping network, the weight mapping network including at least one fully connected layer and an activation function layer, wherein the fully connected layer is configured to perform a linear transformation and shift on the normalized detail evaluation vector, and the activation function layer is configured to use a Sigmoid function to map the output of the fully connected layer to the (0,1) interval to obtain the channel weight vector. w 1, w 2,…, w 256 ].

[0054] S423: Utilize the channel weight vector to analyze the deep low-frequency feature map P D-low (If its number of channels is the same as that of the shallow high-frequency feature map C) L-high Then it is not necessary to perform the above operation of adjusting the number of channels) or the deep low-frequency feature map P after adjusting the number of channels. D-low Channel weighting is performed to obtain the calibration image P. D-low-Calibrate .

[0055] S500: Enhance the details of the image C L-high-Enhance With calibration image P D-low-Calibrate Integration.

[0056] In a simplified embodiment, the detail-enhanced image C L-high-Enhance With calibration image P D-low-Calibrate The fusion method is channel stitching: as described above, the detail enhancement image C L-high-Enhance With calibration image P D-low-Calibrate The spatial dimensions (H, W) are strictly aligned, and the number of channels is adjusted uniformly, that is... , The final merged image F fused =Concat(C L-high-Enhance , P D-low-Calibrate ).

[0057] In a further improved embodiment of the invention, the detail-enhanced image C L-high-Enhance With calibration image P D-low-Calibrate Perform scale-adaptive fusion, such as Figure 4 As shown, adaptive fusion includes the following steps: S510: Enhance the details of the image C L-high-Enhance and calibration image P D-low-CalibrateTemporary channel stitching is performed to obtain a stitched image; the reason for saying "temporary" is that this further embodiment is different from the simple embodiment described above. The channel stitching is not for obtaining the final fused image, but to provide analysis material for the following step S520.

[0058] S520: Analyze the scene-dependent features of the stitched image to determine high-frequency dependence and low-frequency dependence; for example, for an image full of small objects (such as a crowd), the model needs high-frequency details to dominate; for an image with only large objects (such as a car in the distance), the model needs low-frequency semantics to dominate.

[0059] S530: Based on high-frequency dependence and low-frequency dependence, enhance the detail image C L-high-Enhance and calibration image P D-low-Calibrate Assign weights to each: if the high frequency dependency is higher, then the detail-enhanced image C is given a higher weight. L-high-Enhance The larger the assigned weight value, the higher the low-frequency dependence, and the better the calibration image P. D-low-Calibrate The larger the assigned weight value; In one specific embodiment, a pre-trained adaptive weight generation model is used to enhance the detail of image C. L-high-Enhance Assign weight values α And for calibrating image P D-low-Calibrate Assign weight values β : The adaptive weight generation model applies the stitched image Perform global average pooling to compress the spatial features of each channel into a channel-level statistical scalar. , x c ( i , j ) represents the spatial coordinates of the c-th feature channel. i , j The channel feature value at () is in ( H , W The average is calculated across all dimensions to eliminate spatial dimensionality, and the statistical scalars of all channels are combined to obtain the channel feature vectors. ;in, H To stitch images X height, W To stitch images X width, C C Increase the number of channels in the image to enhance detail. C P The number of channels in the calibration image; in embodiments where the detail enhancement image and the calibration image have the same number of channels, , C Enhance the number of channels in an image to improve detail / calibrate the image.

[0060] Each channel-level feature vector is input into an attention network consisting of a neural network and an activation function, wherein the neural network is a fully connected layer or a small MLP network; During training, the attention network updates its parameters through backpropagation of the loss function, configuring it to: identify high-frequency dependency features represented by small objects in the image and low-frequency dependency features represented by large objects in the image; and output weight values ​​for significant high-frequency dependency features. α Tends to be greater than the weight value β For features with significant low-frequency dependence, the output weight value is... β Tends to be greater than the weight value α .

[0061] Specifically, the attention network described above generates a weight vector corresponding to the channel-level feature vector z using a non-linear activation function: ,in, s The weight vector corresponding to the channel-level feature vector z. σ And δ are nonlinear activation functions, W 1 and W 2 is a learnable parameter; in a specific embodiment, σ δ is the Sigmoid function, and δ is the ReLU activation function.

[0062] For the weight vector s The front of the middle C C Each corresponding detail-enhanced image C L-high-Enhance The average of the weighted elements is used as the weight value. α ; for the weight vector s The middle C P One corresponding calibration image P D-low-Calibrate The average of the weighted elements is used as the weight value. β .

[0063] S540: According to the assigned weight values, the detail-enhanced image C... L-high-Enhance With calibration image P D-low-Calibrate Weighted fusion.

[0064] To address the issue that static fusion cannot dynamically adjust its strategy according to the needs of targets at different scales, this embodiment provides a detailed enhancement image C that needs to be fused in the final step. L-high-Enhance and calibration image P D-low-Calibrate Dynamically generate adaptive fusion weights α , βFor detecting shallow features of small targets, higher weights are automatically assigned to high-frequency details; for detecting deep features of large targets, lower-frequency semantics are automatically assigned higher weights, thereby enabling the network to adaptively optimize for multi-scale targets.

[0065] Compared with existing technologies, the bidirectional guided fusion method for image information provided in this invention achieves intelligent complementarity of cross-level features, fundamentally improving fusion efficiency. Existing technologies simply homogenize and mix deep semantic features with shallow detail features, resulting in a blind and inefficient fusion process. This invention introduces a feature frequency band decoupling module, first separating heterogeneous features according to their information attributes (low-frequency semantics / high-frequency details), laying the foundation for refined processing. Subsequently, through the bidirectional guided fusion module, a bidirectional optimization mechanism is established, involving semantic guidance for detail enhancement and detail feedback for semantic calibration. This transforms feature fusion from a passive "mixing" to an active "dialogue," significantly improving the efficiency and accuracy of complementarity between semantic information and spatial details. Secondly, it significantly enhances the detection capability for small-scale targets. This scheme directly addresses the core bottleneck of small target detection. On one hand, the semantically guided detail path can utilize reliable deep semantic information to selectively enhance the details of the corresponding small target regions in the shallow features, while suppressing background noise, effectively alleviating the problem of "high noise" in shallow features. On the other hand, the scale-adaptive aggregation module can automatically allocate more weights to the high-frequency components rich in details in the shallow network path for detecting small targets. The two work together to ensure that the key details of small targets are preserved and enhanced after fusion, thereby effectively reducing the false negative rate and improving the accuracy of localization and recognition. Thirdly, the model's adaptability and robustness are improved. This invention abandons the existing static and fixed fusion method. The scale-adaptive aggregation module enables the network to dynamically adjust the fusion strategy according to the target scale, achieving differentiated optimization for targets of different sizes. At the same time, the bidirectional guidance mechanism itself has adaptability to the input content. This comprehensive adaptive characteristic enables the detection model to exhibit more stable performance and stronger generalization ability when facing diverse and complex scenarios; Fourthly, it maintains modular simplicity, making it easy to integrate and deploy. The entire solution is designed as a plug-and-play bidirectional frequency-domain guided adaptive fusion module, which can directly replace the standard feature pyramid network (such as FPN / PAN) in existing object detection architectures such as YOLO, without changing the main framework of the network or the training process.

[0066] In one embodiment of the present invention, a target detection method is provided, such as... Figure 5 As shown, the detection method includes the following steps: Acquire image information; The acquired image information is fused and enhanced using the bidirectional guided fusion method described above. Based on the fused and enhanced image, such as using a YOLO detection head, the target object can be identified, located, or analyzed.

[0067] like Figure 6 As shown, this embodiment of the invention also provides an image processing system, including the following modules: The feature extraction module is configured to receive image information to be processed and input it into the YOLO model to obtain shallow feature images from the Backbone network of the YOLO model and deep feature images from the Neck network with a PAN structure. An upsampling module is configured to upsample the deep feature image to align it with the shallow feature image. The frequency band decoupling module is configured to perform feature frequency band decoupling on the shallow feature image to obtain the corresponding shallow low-frequency feature map C. L-low and shallow high-frequency feature map C L-high ; and perform feature frequency band decoupling on the upsampled deep feature image to obtain the corresponding deep low-frequency feature map P. D-low and deep high-frequency feature map P D-high ; The detail enhancement guidance module is configured to analyze the deep low-frequency feature map P. D-low The semantic information represented is used to guide the shallow high-frequency feature map C. L-high Enhance details to obtain a detail-enhanced image C L-high-Enhance ; The semantic calibration guidance module is configured to utilize the shallow high-frequency feature map C L-high The weights of each feature channel are evaluated to guide the deep low-frequency feature map P. D-low The calibration semantics are used to obtain the calibration image P. D-low-Calibrate ; The fusion module is configured to fusion the detail-enhanced image C L-high-Enhance With calibration image P D-low-Calibrate The features are then fused to output the final fused feature map.

[0068] This invention also provides a target detection system, including an image acquisition device and an image processing system as described above, wherein the image acquisition device is configured to acquire image information containing a target object; The image processing system is configured to fuse and enhance image information containing the target to be detected; and to identify, locate or analyze the target object based on the fused and enhanced image.

[0069] It should be noted that the image processing system provided in the above embodiments and the image information bidirectional guided fusion method provided in the above embodiments belong to the same inventive concept. Here, the entire contents of the image information bidirectional guided fusion method embodiment are incorporated into this image processing system embodiment by reference, and will not be repeated.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used only 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. Furthermore, 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 limitations, 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.

[0071] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A bidirectional guided fusion method for image information, characterized in that, Includes the following steps: The image information to be processed is input into the YOLO model to obtain the shallow feature image from the Backbone network of the YOLO model and the deep feature image from the Neck network with the PAN structure. The shallow feature image is decoupled by feature frequency band to obtain the corresponding shallow low-frequency feature map C. L-low and shallow high-frequency feature map C L-high ; The deep feature image is upsampled to align with the shallow feature image, and the feature frequency bands of the upsampled deep feature image are decoupled to obtain the corresponding deep low-frequency feature map P. D-low and deep high-frequency feature map P D-high ; Analysis of the deep low-frequency feature map P D-low The semantic information represented is used to guide the shallow high-frequency feature map C. L-high Enhance details to obtain a detail-enhanced image C L-high-Enhance ; and utilize the shallow high-frequency feature map C L-high The weights of each feature channel are evaluated to guide the deep low-frequency feature map P. D-low The calibration semantics are used to obtain the calibration image P. D-low-Calibrate ; The detailed enhanced image C L-high-Enhance With calibration image P D-low-Calibrate Integration.

2. The bidirectional guided fusion method for image information according to claim 1, characterized in that, The detail-enhanced image C L-high-Enhance With calibration image P D-low-Calibrate Scale-adaptive fusion includes: The detailed enhanced image C L-high-Enhance and calibration image P D-low-Calibrate Create temporary channels for stitching to obtain a stitched image; Analyze the scene-dependent features of the stitched image to determine high-frequency and low-frequency dependencies; Based on high-frequency dependence and low-frequency dependence, the detail-enhanced image C is... L-high-Enhance and calibration image P D-low-Calibrate Assign weights to each: if the high frequency dependency is higher, then the detail-enhanced image C is given a higher weight. L-high-Enhance The larger the assigned weight value, the higher the low-frequency dependence, and the better the calibration image P. D-low-Calibrate The larger the assigned weight value; The detail-enhanced image C is then processed according to the assigned weight values. L-high-Enhance With calibration image P D-low-Calibrate Weighted fusion.

3. The bidirectional guided fusion method for image information according to claim 2, characterized in that, Image C is augmented with details using a pre-trained adaptive weight generation model. L-high-Enhance Assign weight values α And for calibrating image P D-low-Calibrate Assign weight values β : The adaptive weight generation model applies the stitched image Global average pooling is performed to compress the spatial features of each channel into a channel-level statistical scalar, resulting in the channel feature vector. ;in, H To stitch images X height, W To stitch images X width, C C Increase the number of channels in the image to enhance detail. C P To calibrate the number of channels in the image; The feature vectors at each channel level are input into an attention network consisting of a neural network and an activation function; During training, the attention network updates its parameters through backpropagation of the loss function, configuring it to: identify high-frequency dependency features represented by small objects in the image and low-frequency dependency features represented by large objects in the image; and output weight values ​​for significant high-frequency dependency features. α Tends to be greater than the weight value β For features with significant low-frequency dependence, the output weight value is... β Tends to be greater than the weight value α .

4. The bidirectional guided fusion method for image information according to claim 3, characterized in that, The attention network generates a weight vector corresponding to the channel-level feature vector z through a non-linear activation function: ,in, s The weight vector corresponding to the channel-level feature vector z. σ And δ are nonlinear activation functions, W 1 and W 2 represents a learnable parameter; For the weight vector s The front of the middle C C Each corresponding detail-enhanced image C L-high-Enhance The average of the weighted elements is used as the weight value. α ; For the weight vector s The middle C P One corresponding calibration image P D-low-Calibrate The average of the weighted elements is used as the weight value. β .

5. The bidirectional guided fusion method for image information according to claim 1, characterized in that, Feature bandwidth decoupling of the feature image is performed on either the shallow feature image or the upsampled deep feature image, in the following manner: The feature image is subjected to low-pass filtering to obtain a low-frequency feature map; and the feature image and the low-frequency feature map are then subjected to a difference operation to obtain a high-frequency feature map. Alternatively, the feature image can be high-pass filtered to obtain a high-frequency feature map; and the feature image and the high-frequency feature map can be differentially analyzed to obtain a low-frequency feature map.

6. The bidirectional guided fusion method for image information according to claim 1, characterized in that, Feature bandwidth decoupling of the feature image is performed on either the shallow feature image or the upsampled deep feature image, in the following manner: The feature image is transformed from the spatial domain to the frequency domain to obtain the spectral information corresponding to the feature image. The spectral information corresponding to the feature image is subjected to low-pass filtering to obtain low-frequency spectral information; and the spectral information corresponding to the feature image and the low-frequency spectral information are differentially processed to obtain high-frequency spectral information; or, the spectral information corresponding to the feature image is subjected to high-pass filtering to obtain high-frequency spectral information; and the spectral information corresponding to the feature image and the high-frequency spectral information are differentially processed to obtain low-frequency spectral information. The low-frequency spectrum information is inversely transformed to obtain a low-frequency feature map in the spatial domain; the high-frequency spectrum information is inversely transformed to obtain a high-frequency feature map in the spatial domain.

7. The bidirectional guided fusion method for image information according to claim 1, characterized in that, Using the deep low-frequency feature map P in the following way D-low The semantic information of the representation guides the shallow high-frequency feature map C L-high Enhanced details: For the deep low-frequency feature map P D-low Semantic confidence analysis was performed to obtain the deep low-frequency feature map P. D-low The semantic confidence of each spatial location; A semantic guidance mask is generated, wherein the element values ​​are mapped to a preset continuous interval by the semantic confidence of each spatial location to represent the semantic saliency of the corresponding spatial location. The semantic guidance mask is used to analyze the shallow high-frequency feature map C. L-high Weighting is performed to obtain the detail-enhanced image C. L-high-Enhance .

8. The bidirectional guided fusion method for image information according to claim 7, characterized in that, Semantic confidence analysis is achieved through a semantic analysis network, which includes sequentially connected convolutional layers and activation function layers. The convolutional layer is used to reduce the number of channels in the deep low-frequency feature map to 1, thus obtaining a single-channel feature map. The activation function layer is configured to use the Sigmoid function to map the feature values ​​of each spatial location in the single-channel feature map to the (0,1) interval, thereby obtaining a semantic confidence map; The method further includes: generating the semantic guidance mask based on the semantic confidence map, including: directly using the semantic confidence map as the semantic guidance mask; or, performing linear scaling or nonlinear transformation on the semantic confidence map to obtain the semantic guidance mask.

9. The bidirectional guided fusion method for image information according to claim 1, characterized in that, The following method is used to obtain shallow high-frequency feature map C L-high Guided deep low-frequency feature map P D-low Calibration semantics: Based on the shallow high-frequency feature map C L-high Calculate the detail evaluation vector that represents the detail richness of each feature channel; Map the detailed evaluation vector to a channel weight vector; The deep low-frequency feature map P is analyzed using the channel weight vector. D-low Channel weighting is performed to obtain the calibration image P. D-low-Calibrate .

10. The bidirectional guided fusion method for image information according to claim 9, characterized in that, The detailed evaluation vector is calculated as follows: The shallow high-frequency feature map C L-high It is a three-dimensional tensor, whose dimensions include channel dimension, height dimension, and width dimension, and its eigenvalues ​​are defined. x c_m ( i , j ) is a three-dimensional tensor with channel index of m The x-axis is i The vertical axis is j scalar element at; For channel index cm The feature channels are analyzed globally along the height and width, and the variance, absolute value, mean, or energy are used to determine the first feature channel. m Feature statistics of each feature channel U m ; The feature statistics of all feature channels are combined in channel index order to obtain the detailed evaluation vector. U detail : U detail =[ U 1, U 2,…, U m ,…, U C ],in, U 1 represents the feature statistics of the first feature channel. U 2 represents the feature statistics of the second feature channel. C This represents the total number of feature channels. U C For the first C Feature statistics for each feature channel.

11. The bidirectional guided fusion method for image information according to claim 9, characterized in that, The detailed evaluation vector is mapped to the channel weight vector in the following manner: The element values ​​of the detail evaluation vector are normalized to obtain the normalized detail evaluation vector; The normalized detail evaluation vector is input into a weight mapping network, which includes at least a fully connected layer and an activation function layer. The fully connected layer is configured to perform a linear transformation and offset on the normalized detail evaluation vector, and the activation function layer is configured to use a sigmoid function to map the output of the fully connected layer to the (0,1) interval to obtain the channel weight vector.

12. The bidirectional guided fusion method for image information according to claim 9, characterized in that, Also includes: The deep low-frequency feature map P is adjusted through 1×1 convolution mapping. D-low The number of channels is equal to that of the shallow high-frequency feature map C. L-high The number of channels; Deep low-frequency feature map P after adjusting the number of channels D-low The calibration image P is obtained by performing a channel-by-channel dot product with the channel weight vector. D-low-Calibrate .

13. The bidirectional guided fusion method for image information according to claim 1, characterized in that, The shallow feature image is the feature image located in the penultimate layer of the Backbone network, and the deep feature image is the feature image output from the last layer of the PAN path aggregation Neck network.

14. A target detection method, characterized in that, Includes the following steps: Acquire image information; The acquired image information is fused and enhanced using the bidirectional guided fusion method for image information as described in any one of claims 1 to 13. Based on the fused and enhanced image, identify, locate, or analyze target objects.

15. An image processing system, characterized in that, Includes the following modules: The feature extraction module is configured to receive image information to be processed and input it into the YOLO model to obtain shallow feature images from the Backbone network of the YOLO model and deep feature images from the Neck network with a PAN structure. An upsampling module is configured to upsample the deep feature image to align it with the shallow feature image. The frequency band decoupling module is configured to perform feature frequency band decoupling on the shallow feature image to obtain the corresponding shallow low-frequency feature map C. L-low and shallow high-frequency feature map C L-high ; and perform feature frequency band decoupling on the upsampled deep feature image to obtain the corresponding deep low-frequency feature map P. D-low and deep high-frequency feature map P D-high ; The detail enhancement guidance module is configured to analyze the deep low-frequency feature map P. D-low The semantic information represented is used to guide the shallow high-frequency feature map C. L-high Enhance details to obtain a detail-enhanced image C L-high-Enhance ; The semantic calibration guidance module is configured to utilize the shallow high-frequency feature map C L-high The weights of each feature channel are evaluated to guide the deep low-frequency feature map P. D-low The calibration semantics are used to obtain the calibration image P. D-low-Calibrate ; The fusion module is configured to fusion the detail-enhanced image C L-high-Enhance With calibration image P D-low-Calibrate The features are then fused to output the final fused feature map.

16. A target detection system, characterized in that, The system includes an image acquisition device and an image processing system as described in claim 15, wherein the image acquisition device is configured to acquire image information containing a target object; The image processing system is configured to fuse and enhance image information containing the target to be detected; and to identify, locate or analyze the target object based on the fused and enhanced image.

Images