A unified object detection method for degraded scenes based on target cues

Abstract

This invention discloses a unified object detection method for degraded scenes based on target cues, belonging to the field of object detection technology. The method includes: constructing a two-stage object detection model based on restored sharp features; wherein the model includes a detection branch, an enhancement branch, and a diffusion denoising module; the enhancement branch is used to extract feature information from the input image as object cues; the diffusion denoising module is used to denoise the object cues to obtain target cues; the detection branch is used to combine the target cues to achieve object detection in degraded scenes; training the model; and using the trained model to achieve object detection in degraded scenes. The technical solution of this invention provides a feasible paradigm for solving the problems of severely insufficient target information in degraded scenes and interference from feature information in various types of complex environments, providing new directions and ideas for further research on degraded scene detection algorithms.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a unified degradation scene target detection method based on target cues. Background Technology

[0002] Object detection refers to the technology of automatically identifying and locating objects in images. Its task is to detect the region and location of a target object in a given image. In recent years, deep learning-based object detection algorithms have made great progress and have branched into two important directions: methods using only convolutional neural networks (CNNs) and methods based on Transformers. CNN-only networks can be further divided into two-stage detection frameworks (including a preprocessing step for generating object proposals) and single-stage detection frameworks.

[0003] Two-stage object detection algorithms are object detection methods based on Region Proposal Networks (RPNs). They consist of two steps: generating candidate object regions (Proposals) and processing these regions to predict the object's class and bounding box. In the first step, the RPN uses a sliding window method to extract regions from the image. Each proposed region is defined as a rectangle containing the sample object, and these regions are mapped onto a smaller, fixed-size feature map using a CNN, allowing for shared computation. In the second step, the detection model uses these proposed regions as input and performs classification and regression to determine the predicted object's bounding box and class. The main advantage of this method is its high detection accuracy, but its computational complexity is relatively high.

[0004] One-stage object detection algorithms do not require an RPN stage. They directly generate the object's class probability and location coordinates, and output the detection result directly through a single forward pass, thus offering faster detection speeds. YOLO treats object detection as a regression problem involving spatially separated bounding boxes and associated class probabilities. The entire detection process is a single network that predicts bounding boxes and class probabilities from the complete image in a single evaluation, allowing for end-to-end optimization of detection performance.

[0005] Transformer-based object detection algorithms have become a highly popular research area in recent years. Their main advantage is that the Transformer model can replace the traditional R-CNN module, eliminating the need for general object detection procedures such as anchor boxes and non-maximum suppression (NMS). Instead, the entire image can be directly encoded, and all object detection box predictions are treated as a standard unordered set prediction problem. Furthermore, a self-attention mechanism is used to capture the dependencies between objects.

[0006] While general-purpose object detection networks perform excellently in standard scenarios, directly applying them to degraded scenes often presents numerous challenges. Degraded scenes include complex conditions such as low light, fog, rain, snow, motion blur, occlusion, and low resolution. These factors significantly degrade image quality, leading to blurred or lost object features, thus impacting detection performance. Issues such as degraded image quality, unclear object features, reduced multi-scale detection performance, and insufficient model robustness all contribute to a significant decrease in detection accuracy.

[0007] Object detection in degraded scenes is an important research direction in computer vision, aiming to solve the problem of accurately identifying and locating objects under conditions of degraded image quality, complex environments, or unclear object features. Existing methods have attempted to address these problems from various aspects, but most methods still involve pixel-level image enhancement, introducing a large amount of computational redundancy, and have not made them entirely beneficial for detection.

[0008] (1) Direct Training Method: The target detector is directly retrained using degraded images. Its advantages lie in end-to-end optimization and avoidance of cascaded errors, but its disadvantages are significant. First, it is highly dependent on data, requiring a large amount of labeled data for degraded scenes. However, in real-world scenarios, there are diverse types of degradation, making data acquisition and labeling extremely costly. Second, its generalization ability is limited; models trained for specific degradation types are difficult to adapt to other degraded scenes (such as the simultaneous presence of haze and low light). Furthermore, due to the low quality of the original images in degraded scenes, direct detection may struggle to extract effective features, leading to performance bottlenecks.

[0009] (2) Cascaded Training Method: Image enhancement and detection are cascaded in stages. The degraded image is enhanced first and then input into a pre-trained detector. The advantage of this method is its modular design and flexible replacement, but its disadvantages are more prominent. First, the staged nature of enhancement and detection results in severe computational redundancy, making it difficult to meet real-time requirements. Second, the performance of the enhancement algorithm directly affects the detection results. If enhancement fails (e.g., excessive denoising leads to target loss), the detector is difficult to remedy, leading to error accumulation. In addition, image enhancement is visual quality-oriented (e.g., PSNR, SSIM), while detection needs to retain key target features. The two objectives may conflict, further limiting detection performance.

[0010] (3) Joint Training Method: Theoretically, this method, which involves joint training of the augmentation and detection modules end-to-end, holds the most potential, but its drawbacks cannot be ignored. First, it has high design complexity, requiring careful design of the network structure and loss function, making training difficult. Second, joint models typically have a large number of parameters, demanding high data and computing resources, making them difficult to deploy in resource-constrained scenarios. Furthermore, the coupling between augmentation and detection may make the intermediate processes difficult to analyze, reducing the model's interpretability. Although joint training can theoretically achieve a globally optimal solution, its implementation complexity and high data and computing power requirements limit its widespread adoption in practical applications.

[0011] In summary, existing target detection algorithms have low detection accuracy in degraded scenarios and cannot meet the requirements. Summary of the Invention

[0012] This invention provides a unified target detection method for degraded scenes based on target cues, in order to solve the technical problem that existing target detection algorithms have low detection accuracy in degraded scenes and cannot meet the requirements.

[0013] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0014] On the one hand, the present invention provides a unified degradation scene target detection method based on target cues, including:

[0015] A two-stage target detection model based on restored clear features is constructed. The model includes a detection branch, an enhancement branch, and a diffusion denoising module. The enhancement branch is used to extract feature information from the input image as object cues. The diffusion denoising module is used to denoise the object cues to obtain target cues. The detection branch is used to combine the target cues to achieve target detection in degraded scenes.

[0016] Training a two-stage object detection model based on restored clear features;

[0017] A pre-trained two-stage target detection model based on restored clarity features is used to achieve target detection in degraded scenes.

[0018] Furthermore, the two-stage target detection model based on restored clear features adopts a two-stage decoupled training mechanism;

[0019] In the first stage of training, the detection branch and the enhancement branch are trained; wherein, the enhancement branch takes a clear image as input, extracts feature information from the clear image as object cues, and inputs the object cues into the detection branch; the detection branch takes a degraded image and the object cues output by the enhancement branch as input, performs target detection on the degraded image, and optimizes the first stage of training by minimizing the detection loss;

[0020] In the second phase of training, the enhancement branch is frozen, the diffusion denoising module is trained, and the detection branch is fine-tuned. The diffusion denoising module takes the object cues output by the enhancement branch as input, and its output is input to the detection branch. It is trained using consistency loss as intermediate supervision while maintaining the detection loss constraint, so as to achieve joint training of the diffusion denoising module and the enhancement branch.

[0021] Furthermore, the diffusion denoising module constructs a denoising process based on a diffusion model, and uses the diffusion model to denoise the object cues output by the enhancement branch to obtain the target cues.

[0022] Furthermore, the inference process of the two-stage target detection model based on restored sharp features includes:

[0023] Using a degraded image as input, the degraded image is mapped to a pseudo-sharp domain to obtain an enhanced image;

[0024] The enhanced image is input into the enhancement branch to obtain the object cue Z corresponding to the enhanced image. e ;

[0025] Z e Input the diffusion denoising module, with Z e As a conditional input to the diffusion model, it controls the diffusion model to gradually remove the degradation effects in the iterative denoising process, and obtain the feature estimate of the clear image as a target cue;

[0026] The target clues and the degraded image are input into the detection branch to achieve target detection in the degraded image.

[0027] Furthermore, the detection branch includes: a detection backbone network, a feature fusion module, and a detection head; among which,

[0028] The detection backbone network performs spatial downsampling on the input degraded image through multiple stages of stacked convolutional layers to obtain multi-scale feature maps, which serve as feature information of the degraded image.

[0029] The feature fusion module fuses the feature information of the degraded image with the target clues to obtain fused features;

[0030] The detection head achieves target localization and classification based on fused features, through anchor frame mechanism or key point prediction.

[0031] Furthermore, the feature fusion module fuses the feature information of the degraded image with the target cues to obtain the fused features. The process includes: outputting the feature map F in the nth stage of the detection backbone network. n Afterwards, F n Separate into F along the channel α and Fβ , of which F α The target cues are concatenated along the channels, and then fused and convolved using the C3 module in the YOLOv5 detector, with the number of channels adjusted, to obtain enhanced features; these enhanced features are then combined with the F... β The features are reassembled to obtain the final fused features, which serve as the input for the next stage in the detection backbone network; where n is a preset value.

[0032] Furthermore, the detection backbone network deploys feature fusion modules at the P3, P4, and P5 levels of the feature pyramid.

[0033] Furthermore, during the corresponding level of processing, the target clues will be bilinearly interpolated using the corresponding downsampling rate.

[0034] Furthermore, the enhanced branch extracts object cues corresponding to the input image by stacking convolutional layers.

[0035] In another aspect, the present invention also provides an electronic device comprising a processor and a memory; wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described method.

[0036] In another aspect, the present invention also provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the above method.

[0037] The beneficial effects of the technical solution provided by this invention include at least the following:

[0038] The technical solution of this invention employs a two-stage collaborative training mechanism, decoupling the detection task from image restoration during training, thus optimizing the conflict between image restoration and target detection in traditional end-to-end training. A hierarchical feature fusion module is proposed, using target cues to adaptively compensate for degraded features at levels P3-P5 of the feature pyramid. Furthermore, target cue extraction is modeled as a latent space conditional diffusion process, eliminating dependence on external image enhancement and achieving end-to-end mapping from noisy space to clear latent space. This provides a feasible paradigm for addressing the severe lack of target information in degraded scenes and the interference of feature information from various types of complex environments, offering new directions and ideas for further research on degraded scene detection algorithms, and possessing significant application prospects. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the unified degradation scene target detection method based on target cues provided in the embodiments of the present invention;

[0041] Figure 2 This is a framework diagram of a two-stage target detection model based on restored clear features provided in an embodiment of the present invention;

[0042] Figure 3 This is a system block diagram of the electronic device provided in the embodiments of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0044] First, it should be noted that in the embodiments of the present invention, the words "exemplarily," "for example," etc., are used to indicate that they are examples, illustrations, or descriptions. Any embodiment or design scheme described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of the term "exemplarily" is intended to present the concept in a specific manner. Furthermore, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either one or the other.

[0045] First Embodiment

[0046] This embodiment provides a unified object detection method for degraded scenes based on target cues. It employs a two-stage decoupled training mechanism: the first stage utilizes clear image supervision to construct a latent space for target cues; the second stage achieves degradation-robust latent space mapping through a conditional diffusion model. Furthermore, it constructs a cue-driven feature compensation: dynamically adjusting the cross-level fusion of feature pyramids (P3-P5) to achieve adaptive degradation compensation that preserves shallow details while maintaining semantic consistency at deeper levels. This method can be implemented by an electronic device, such as a terminal or a server. The execution flow of this method is as follows: Figure 1 As shown, it includes the following steps:

[0047] S1, Construct a two-stage target detection model based on restored clear features; wherein, the model includes a detection branch, an enhancement branch, and a diffusion denoising module; the enhancement branch is used to extract feature information of the input image as object cues; the diffusion denoising module is used to denoise the object cues to obtain target cues; the detection branch is used to combine the target cues to achieve target detection in degraded scenes;

[0048] S2, train the two-stage target detection model based on restored clear features;

[0049] S3 utilizes a trained two-stage target detection model to achieve target detection in degraded scenes.

[0050] It's important to note that current object detection algorithms for degraded scenes generally employ a sequential processing paradigm of "enhancement before detection." This means that low-quality input is preprocessed using pixel-level image enhancement techniques, and the enhanced result is then used as the sole input to the detection network. This paradigm suffers from several bottlenecks: First, the enhancement network is designed in isolation from the detection task, lacking a multi-task collaborative optimization mechanism. Although the PSNR / SSIM scores of the enhanced image are significantly improved, its feature distribution deviates from the detector's preferences, failing to explicitly model the high-level semantic features required for the detection task. This makes it difficult for the detection network to effectively utilize the fine-grained texture information extracted during the enhancement stage. Second, mainstream enhancement models over-rely on specific degradation physics priors, leading to a sharp performance drop in complex scenes with heterogeneous degradation factors (such as haze, motion blur, and noise). Furthermore, even some joint optimization schemes remain confined to the pixel-level enhancement framework, failing to address redundant computation and ineffective feature transfer issues. This dual limitation in feature adaptability and scene generalization severely restricts the practical value of the algorithms in real-world open environments.

[0051] To address the aforementioned issues, this embodiment proposes a two-stage object detection model based on restored sharp features, employing a two-stage collaborative training mechanism. In the first stage, during training, a new sharp feature generation branch is added, which extracts sharp features from the sharp image I using a salient detail extraction module. c Extracting clear features Z c These features contain information that is not easily noticeable or even missing in degraded images. Subsequently, using a feature embedding module and downsampling operations, we... c The feature maps at three scales are flexibly injected into the object detection backbone network and fully integrated with the original degraded features, thereby improving detection performance. In the second stage of training, we restore the feature Z... c This removes the model's dependence on external information while avoiding computational redundancy caused by pixel-level image restoration. Because from the degraded image I... d Z recovers clear image information c The process is extremely similar to the gradual denoising process at the feature layer level in the Latent Space Diffusion Model (LDM). Therefore, we only input the degraded image I. d In this case, a conditional diffusion model is used to estimate the sharp feature Z. c Specifically, we also use the salient detail extraction module to extract details from I. d Extracting feature Z dThis information is then used as a condition input into the diffusion model, controlling the model to gradually remove degradation effects and restore detailed information during iterative denoising, ultimately yielding high-quality, clear feature estimates. Then, similar to the first stage, the restored features are injected into the target detection backbone network for joint fine-tuning, resulting in a two-stage detection network based on the restored features. Furthermore, our first-stage network, combined with the underwater augmentation network, provides another complete solution.

[0052] The overall architecture of this two-stage object detection model based on restored clear features is as follows: Figure 2 As shown, the algorithm includes a detection branch, an enhancement branch, and a diffusion denoising module. The enhancement branch takes a sharp image as input and extracts sharp object cues through stacked convolutional layers. The detection branch uses a feature fusion module (FFM) to integrate object cues with the original detection backbone to improve object detection in visually degraded scenes. To extract object cues directly from degraded images, we leverage the powerful data estimation capabilities of the diffusion model, using degraded object cues as conditions, by... c Replace with A unified inference pipeline has been implemented that does not rely on external enhancement methods or degenerate specific priors. The model's workflow can be summarized as follows:

[0053] Given a degraded image I d As input, our goal is to... d Target object detection is performed within the target detection branch, which includes a detection backbone network and a detection head. To compensate for I... d Compared to a clear image I c To address the degradation effect, we propose extracting cues helpful for object detection, termed target cues Z, through image enhancement branching and diffusion modules, and then using I during the training phase. c As a guide. Specifically, such as Figure 1 As shown, the image enhancement branch produces a clearer image I. c (or enhanced image I) e As input, output Z c (or Z) e In order to achieve results without relying on I c (or I) e In the case of obtaining target clues, our diffusion model module uses I d As input, a denoising process based on a diffusion model is constructed, according to Z... d Directly estimate Z c Finally, the estimated target cues (i.e., ...) are fused together using the feature fusion module. This is integrated into the object detection branch to enhance object detection performance in visually degraded scenarios. The core technologies of the model will then be described in detail module by module.

[0054] 1. Detection Branch

[0055] This module proposes an adaptive feature enhancement method for degraded scenarios. Through a multi-scale target cue-guided feature reconstruction mechanism, it effectively improves the robustness of target detection under complex imaging conditions. It should be noted that traditional target detection frameworks typically employ a cascaded architecture of "backbone network-detector head," where the backbone network progressively performs spatial downsampling through N stacked convolutional layers, ultimately outputting multi-scale feature maps. The detector head then uses these feature maps to achieve target localization and classification through anchor box mechanisms or keypoint prediction. Since different detection frameworks have different "backbone network-detector head" designs, in this implementation, we use Faster R-CNN with MobileNetV3-large as the backbone network as the baseline. We propose a target cue-guided feature enhancement paradigm. This paradigm does not rely on image quality enhancement preprocessing but instead establishes a degradation-robust feature compensation mechanism by constructing a target cue matrix in the latent space. The acquisition of the target cue matrix will be detailed in the next two branches. Specifically, the feature map F output in the nth stage of the backbone network... n Next, a target cue-guided feature fusion module (FFM) is introduced. This module employs a two-branch processing mechanism: first, the F... n Separate along the channel into F α and F β F α The Z-axis is concatenated along the channel, and then the C3 module, commonly used in YOLOv5 detectors, is used as the main fusion convolution operation with the number of channels adjusted to obtain enhanced features. Finally, the enhanced features are combined with F... β The features are reassembled to obtain the final fused features, which serve as input for the next stage in the backbone network. To ensure consistent enhancement of multi-scale features, this method deploys FFM modules at three key levels of the feature pyramid: P3 (8× downsampling), P4 (16× downsampling), and P5 (32× downsampling). n During layer processing, target cues are bilinearly interpolated using the corresponding downsampling rate, resulting in scale-aware feature enhancement. This hierarchical design enables the network to: preserve target detail features in shallow layers; maintain semantic consistency in deeper layers; and achieve feature compensation through cross-scale information flow. Notably, due to its versatility, the design can be transferred to any framework based on convolutional networks.

[0056] 2. Enhance branches:

[0057] During the training phase, clear images can be directly acquired as supervision signals. At this point, feature decoupling can be performed using a salient detail extraction (SDE) module to obtain a high-confidence target cue matrix Z. cIt encodes detection-sensitive features such as target edges and textures. During the inference phase, to address the constraint that sharp images are unavailable in real-world scenes, we design a degradation-robust augmentation-detection closed-loop architecture. We use image augmentation methods as a preprocessing step to map the degraded image to the pseudo-sharp domain to obtain the augmented image I. e This image is used to replace the clear image as input to the SDE to complete the entire inference process. Since there is no single method that can enhance the image quality of any scene, the choice of model is usually determined by the specific degradation scenario, such as DEANet for foggy scenes and GLARE for low-light scenes. The SDE module design is as follows: Figure 2 As shown, it consists of several convolutional kernels of different sizes and lengths stacked together, and finally a 1×1 convolution is used to compress the channels to 1. However, this method has a serious limitation: during inference, it must rely on external input, making it tied to image enhancement methods.

[0058] 3. Diffusion denoising module:

[0059] To construct an end-to-end unified inference framework independent of external image enhancement and achieve adaptive extraction of target cues in degraded scenes, we propose a cross-domain modeling method based on latent space feature restoration. Pixel-domain-based target cue extraction schemes are limited by the prior physical imaging differences of different degradation types in the pixel domain, making it difficult to establish a universal degradation-sharpness mapping relationship in the pixel domain for direct and effective target cue extraction. Therefore, we turn to the latent representation space, modeling target cue restoration as a conditional denoising process, and establishing a mapping relationship from the latent space of the sharp image to the space of the degraded image.

[0060] Forward Process: The forward process of the diffusion model is a gradual addition of noise. It involves progressively perturbing the input using a parameterized Markov chain, i.e., representing the Z-value of the latent space of the sharp image. c By gradually adding Gaussian noise, the model is mapped to a noise space. During this gradual noise addition process, different types of degradation become indistinguishable due to the noise, which is more conducive to establishing a unified framework.

[0061] The reverse process: The reverse process of the diffusion model is a stepwise denoising process. It achieves a mapping from the noisy space to the sharp latent space through progressive denoising. This process starts with random Gaussian noise and then proceeds through the degenerate target cue Z. d Guided by this approach, noise interference is gradually removed through iterative refinement, ultimately reconstructing a clear latent space representation of the image. This achieves a mapping from the noisy space to the clear latent space. Specifically, within the framework of the conditional diffusion model, the inverse denoising process uses randomly sampled Gaussian noise as the initial state. In each iteration, the degenerate target cue Z generated by SDE is used... dAs a conditional signal, it is embedded into the U-Net backbone network through an attention mechanism, guiding the network to accurately separate noise components from essential features. After T rounds of progressive optimization, the final output is a reconstruction result aligned with the latent space distribution of the clear image. Thanks to the efficient and compact representation of the latent space, and the inherent similarity between the conditional matrix and the target representation, convergence to the optimal solution is achieved in just 5 iterations. Furthermore, the denoising network employs a streamlined U-Net architecture, using fewer intermediate layer channels and stacked residual blocks, making our diffusion model more lightweight.

[0062] 4. Model training and inference:

[0063] During training, directly extracting target cues from degraded images without intermediate guidance is challenging. Therefore, we employ a two-stage training strategy to extract target cues from degraded images. In the first stage, we focus on training the detection and enhancement branches, implementing the target cues as feature maps of a latent spatial representation driven by the object detection task, i.e., extracting sharp target cues Z from sharp images. c Degraded and clear images are used as input to jointly train the detection backbone and detection head. The detection branch accepts Z-axis input. c As cues for object detection, the first stage of training is optimized by minimizing the detection loss. In the second stage of training, the focus is on training the diffusion model and fine-tuning the detection model. Specifically, only degraded images are used as input, degraded object cues are extracted as conditional inputs to the diffusion model, the enhancement branch is frozen, and its output Z-value is used. c As a guide, consistency loss is used as an intermediate supervision to train the diffusion model, while maintaining the detection loss constraint to preserve the compatibility of the feature space with the detection task.

[0064] During inference, there are two modes: 1) The detection branch uses the image extracted from the augmented image as additional input, corresponding to the augmentation branch mentioned above. It's worth noting that since the augmentation model and the detection branch are not jointly fine-tuned, this method is considered a cascaded method, called D4Det-Sep. 2) The detection branch uses the image restored by DM... As an additional input, this corresponds to the diffusion model denoising process described above, and is our final unified method, called D4Det.

[0065] In summary, this embodiment provides a unified degraded scene object detection method based on target cues. Its core features include: a two-stage collaborative training mechanism: decoupling the detection task from image restoration during training, optimizing the conflict between image restoration and object detection in traditional end-to-end training; a multi-scale target cue fusion architecture: proposing a hierarchical feature fusion module, using target cues to adaptively compensate for degraded features at the P3-P5 levels of the feature pyramid; and a latent space conditional denoising paradigm: modeling target cue extraction as a latent space conditional diffusion process, eliminating dependence on external image enhancement and achieving end-to-end mapping from noisy space to clear latent space. This unified degraded scene object detection method based on target cues has the following advantages compared to existing technologies:

[0066] (1) Comparison with physical prior-based augmentation methods: Existing methods typically design dedicated networks for single degradation scenarios, such as atmospheric scattering models for foggy weather and Retinex theory for low-light conditions. While these methods can achieve excellent results in specific scenarios, their strong scene dependence severely restricts their cross-scenario generalization ability. Therefore, this paper proposes a target-cue-driven feature space decoupling framework that maps images to an implicit feature space and adaptively learns shared degradation representations across scenarios, avoiding dependence on explicit physical assumptions. At the same time, it forces the network to retain only salient features common to cross scenarios while suppressing task-irrelevant redundant information and strengthening general detailed representations.

[0067] (2) Comparison with other pixel-level image enhancement methods: Pixel-level enhancement methods focus on low-level visual quality optimization in the pixel domain, such as contrast enhancement and noise suppression, but neglect the synergistic enhancement of high-level semantic information and detector-sensitive features. Therefore, we construct a mapping from the degraded image to the latent feature space, modeling feature enhancement as a progressive denoising process in the latent space. That is, with degraded features and random noise as the initial state, we iteratively approximate the detection-friendly ideal feature distribution through back-diffusion. This mechanism dynamically decouples the degraded interference from the essential features of the target through a noise prediction network, so that the enhanced feature map not only has visual rationality but also retains information conducive to detection.

[0068] The effectiveness of the method in this embodiment will now be verified.

[0069] Table I shows the performance comparison of our method with other state-of-the-art methods on the foggy VOC-F and RTTS datasets. We selected Faster-RCNN with the MobilenetV3-large backbone as the benchmark. *In the table, "Baseline" indicates that the training data consists of sharp images. Compared to other cascaded methods, D4Det-Sep significantly outperforms the corresponding cascaded methods under the same augmentation conditions. On the synthetic VOC-F dataset, while using DEANet as the augmentation network, the cascaded methods resulted in a 1.02% decrease in the map performance, while our method improved it by 2.21%. D4Det achieved further improvements compared to D4Det-Sep. Meanwhile, on the real foggy RTTS dataset, its performance improved by 6.67% compared to the baseline and by 4.13% compared to JE-YOLO, the best among the joint methods. This indicates that, without relying on external image augmentation networks, using only degraded images as input, a diffusion model can refine the target cues more effectively. Experimental results validate the effectiveness of our proposed method.

[0070] Table I

[0071] Performance comparison with existing methods on foggy datasets

[0072]

[0073] The performance of our method on the low-light VOC-D and ExDark datasets, compared with other state-of-the-art methods, is shown in Table II. Table II reveals that D4Det-Sep, optimized for target cues, outperforms all other direct and cascaded methods, demonstrating a significant detection gain on the ExDark dataset. Our D4Det also exhibits the highest mAP performance on both synthetic and real datasets.

[0074] Table II

[0075] Performance comparison with existing methods on low-light datasets

[0076]

[0077] Furthermore, we extended the case to a more general scenario where no corresponding sharp image exists during training. Therefore, we preprocessed the degraded images using augmentation methods and used the augmented images instead of the sharp ones to complete the entire training process. As shown in Table III, our D4Det still demonstrates superior performance compared to other methods on the underwater datasets URPC and DUO.

[0078] Table III

[0079] Performance comparison with existing methods on underwater datasets

[0080]

[0081] Extensive experiments conducted under foggy, low-light, and underwater conditions demonstrate that the proposed method outperforms other existing methods. The algorithm presented in this invention provides a feasible paradigm for addressing the severe lack of target information in degraded scenes and the interference of various complex environmental features. Since it does not involve any prior physical models, this algorithm can be extended to vehicle detection in degraded scenes such as rain and snow, in addition to the scenarios described above, providing new directions and ideas for further research on degraded scene detection algorithms.

[0082] Second Embodiment

[0083] This embodiment provides an electronic device, such as... Figure 3 As shown, the electronic device includes a processor and a memory; wherein the processor and the memory can be connected via a communication bus; the memory stores at least one instruction, which is loaded and executed by the processor to implement the method of the first embodiment described above. Furthermore, the electronic device may also include a transceiver, the processor and the transceiver can be connected via a communication bus, and the transceiver is used to communicate with other devices.

[0084] Below, in conjunction with Figure 3 A detailed introduction to each component of this electronic device is provided below:

[0085] The processor is the control center of the electronic device. The electronic device may include multiple processors, each of which can be a single-core processor (single-CPU) or a multi-core processor (multi-CPU). The term "processor" can refer to a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), other general-purpose processors, application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), one or more field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor can perform various functions of the electronic device by running or executing software programs stored in memory and by calling data stored in memory.

[0086] In a specific implementation, as one example, the processor may include one or more CPUs, for example... Figure 3 CPU0 and CPU1 shown are, of course, merely illustrative examples.

[0087] The memory is used to store the software program that executes the solution of the present invention, and the processor controls its execution. For specific implementation methods, please refer to the above method embodiments, which will not be repeated here.

[0088] Optionally, the memory may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory may be integrated with the processor or may exist independently, and may be accessed through the interface circuit of the electronic device (…). Figure 3 (Not shown in the image) is coupled to the processor; however, this embodiment of the invention does not impose specific limitations on this.

[0089] The transceiver may include a receiver and a transmitter. Figure 3 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function. The transceiver can be integrated with the processor or exist independently, and is connected through the interface circuit of the electronic device (…). Figure 3 (Not shown in the image) is coupled to the processor, and this embodiment of the invention does not specifically limit this.

[0090] In addition, it should be noted that, Figure 3 The structure of the electronic device shown is not intended to limit the device. Actual devices may include more or fewer components than shown, or combine certain components, or have different component arrangements. Furthermore, the technical effects achieved by this electronic device when performing the method of the first embodiment described above can be referenced to the technical effects described in the first embodiment; therefore, they will not be repeated here.

[0091] Third embodiment

[0092] This embodiment provides a computer-readable storage medium storing at least one instruction, which is loaded and executed by a processor to implement the method of the first embodiment described above. The computer-readable storage medium may be a ROM, random access memory, CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc. The instruction stored therein can be loaded and executed by a processor in a terminal.

[0093] Furthermore, it should be noted that the present invention can be provided as a method, apparatus, or computer program product. Therefore, embodiments of the present invention can take the form of a completely or partially hardware embodiment, a completely or partially software embodiment, or an embodiment combining software and hardware aspects. Moreover, when implemented in software, embodiments of the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer program are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any usable medium accessible to a computer or a data storage device such as a server or data center containing one or more sets of usable media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive (SSD).

[0094] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 a process or multiple processes and / or boxes Figure 1The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0096] It should also 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. The terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device 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 terminal device. 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 terminal device that includes said element. Furthermore, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone, where A and B can be singular or plural. Additionally, the character " / " in this text generally indicates an "or" relationship between the preceding and following objects, but it can also indicate an "AND / OR" relationship. Please refer to the context for specific interpretations. "At least one" refers to one or more items, while "more than" refers to two or more items. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c can be represented as: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0097] Furthermore, it is understood that in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0098] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0099] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of functional modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms. Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, i.e., they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs. Additionally, the functional units in the various embodiments of this invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0100] If the method is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments of the present invention have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make several improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A unified degraded scene target detection method based on target cues, characterized in that, include: A two-stage target detection model based on restored clear features is constructed. The model includes a detection branch, an enhancement branch, and a diffusion denoising module. The enhancement branch is used to extract feature information from the input image as object cues. The diffusion denoising module is used to denoise the object cues to obtain target cues. The detection branch is used to combine the target cues to achieve target detection in degraded scenes. Training a two-stage object detection model based on restored clear features; A pre-trained two-stage target detection model based on restored clarity features is used to achieve target detection in degraded scenes; The two-stage target detection model based on restored clear features adopts a two-stage decoupled training mechanism; In the first stage of training, the detection branch and the enhancement branch are trained; wherein, the enhancement branch takes a clear image as input, extracts feature information from the clear image as object cues, and inputs the object cues into the detection branch; the detection branch takes a degraded image and the object cues output by the enhancement branch as input, performs target detection on the degraded image, and optimizes the first stage of training by minimizing the detection loss; In the second stage of training, the enhancement branch is frozen, the diffusion denoising module is trained, and the detection branch is fine-tuned. The diffusion denoising module takes the object cues output by the enhancement branch as input, and its output is input to the detection branch. It is trained using consistency loss as intermediate supervision while maintaining the detection loss constraint, so as to achieve joint training of the diffusion denoising module and the enhancement branch. The diffusion denoising module constructs a denoising process based on a diffusion model, and uses the diffusion model to denoise the object cues output by the enhancement branch to obtain the target cues. The reasoning process of the two-stage target detection model based on restored clear features includes: Using a degraded image as input, the degraded image is mapped to a pseudo-sharp domain to obtain an enhanced image; The enhanced image is input into the enhancement branch to obtain the object cues corresponding to the enhanced image. ; Will Input the diffusion denoising module, to As a conditional input to the diffusion model, it controls the diffusion model to gradually remove the degradation effects in the iterative denoising process, and obtain the feature estimate of the clear image as a target cue; The target clues and the degraded image are input into the detection branch to achieve target detection in the degraded image.

2. The unified degradation scene target detection method based on target cues as described in claim 1, characterized in that, The detection branch includes: a detection backbone network, a feature fusion module, and a detection head; wherein... The detection backbone network performs spatial downsampling on the input degraded image through multiple stages of stacked convolutional layers to obtain multi-scale feature maps, which serve as feature information of the degraded image. The feature fusion module fuses the feature information of the degraded image with the target clues to obtain fused features; The detection head achieves target localization and classification based on fused features, through anchor frame mechanism or key point prediction.

3. The unified degradation scene target detection method based on target cues as described in claim 2, characterized in that, The feature fusion module fuses the feature information of the degraded image with the target cues to obtain the fused features. The process includes: outputting a feature map in the nth stage of the detection backbone network. Afterwards, Separated along the channel and ,in, The target cues are concatenated along the channels, and then fused and convolved using the C3 module in the YOLOv5 detector, with the number of channels adjusted, to obtain enhanced features; these enhanced features are then combined with... The features are reassembled to obtain the final fused features, which serve as the input for the next stage in the detection backbone network; where n is a preset value.

4. The unified degradation scene target detection method based on target cues as described in claim 3, characterized in that, The detection backbone network deploys feature fusion modules at the P3, P4, and P5 levels of the feature pyramid.

5. The unified degradation scene target detection method based on target cues as described in claim 4, characterized in that, During processing at the appropriate level, the target clues will be bilinearly interpolated using the corresponding downsampling rate.

6. The unified degradation scene target detection method based on target cues as described in claim 1, characterized in that, The enhancement branch extracts object cues corresponding to the input image by stacking convolutional layers.

Images