Target detection method and device for deblurred image

Abstract

The invention discloses a target detection method and device for a deblurred image, and belongs to the field of image recognition. The method comprises the following steps: acquiring a deblurred target detection data set and a YOLOv8n baseline model; performing structure optimization on the YOLOv8n baseline model to obtain an optimized detection model; training and optimizing the detection model by adopting the target detection data set, and generating a deblurred image detection model; and when a to-be-detected deblurred image is received, the features of the deblurred image are extracted through the deblurred image detection model, and the target position and category are predicted.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and more specifically to a method and device for target detection in deblurred images. Background Technology

[0002] Object detection, a core task in computer vision, is complex in many aspects, including multi-object tracking, occlusion handling, adaptation to pose diversity, and scaling. Convolutional Neural Networks (CNNs), in particular, have significant applications in image processing, video analysis, and medical imaging. Therefore, continuous and in-depth research into object detection technology not only aligns with current academic trends but also meets society's demand for intelligent solutions.

[0003] However, existing blurry image detection processes often face problems such as blurred target contours, severe noise interference, and damage to edge and detail information.

[0004] Therefore, how to provide a target detection method and device for deblurred images is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a method and apparatus for target detection in deblurred images.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A target detection method for deblurred images includes:

[0008] Obtain the deblurred object detection dataset and the YOLOv8n baseline model;

[0009] The YOLOv8n baseline model is structurally optimized to obtain an optimized detection model;

[0010] The optimized detection model is trained using the target detection dataset to generate a deblurred image detection model;

[0011] When a deblurred image to be detected is received, the features of the deblurred image are extracted and the target location and category are predicted by the deblurred image detection model.

[0012] Furthermore, structural optimization is performed on the YOLOv8n baseline model, including:

[0013] The neck network of the YOLOv8n baseline model is replaced with a high-level feature and low-level feature block fusion module HLBFM, which is used to collect and fuse feature information from all levels and redistribute it to different levels.

[0014] The C2f module in the backbone network of the YOLOv8n baseline model is replaced with the C2f_CReToNeXt module, which is used to extract multi-scale features by connecting multiple sub-networks in parallel.

[0015] Furthermore, the high-level feature and low-level feature block fusion module HLBFM includes:

[0016] The Low-GD branch, which collects and distributes low-level features, is used for downsampling and feature fusion of input features.

[0017] The High-GD branch, which focuses on advanced feature collection and distribution, is used for global information extraction and fusion of input features.

[0018] Furthermore, the Low-GD branch for low-level feature collection and distribution includes: a Low-FAM module, a Low-IFM module, and a first information injection module;

[0019] The Low-FAM low-level feature alignment module downsamples the input features using average pooling.

[0020] The Low-IFM low-level information fusion module achieves feature fusion through reparameterizable convolutional blocks (RepBlock) and segmentation operations.

[0021] The first information injection module is used to perform attention fusion between local feature information and global injected information.

[0022] Furthermore, the High-GD branch of advanced feature collection and distribution includes: a High-FAM module, a High-IFM module, and a second information injection module.

[0023] The High-FAM module reduces the input features to a uniform size through average pooling.

[0024] The High-IFM (High-Information Fusion) module extracts and fuses global features through the Transformer block.

[0025] The second information injection module is used to perform attention fusion between local feature information and global injected information.

[0026] Furthermore, both the first information injection module and the second information injection module include:

[0027] The globally injected information is processed through two independent convolutional layers to obtain the first feature and the second feature;

[0028] The local feature information is processed through a convolutional layer to obtain the third feature;

[0029] The first feature, the second feature, and the third feature are fused using an attention mechanism.

[0030] Furthermore, the C2f_CReToNeXt module includes:

[0031] The branching processing mechanism divides the input features into two independent branches, and performs differential convolution operations on each branch.

[0032] The feature merging module is used to merge the outputs of two branches.

[0033] Furthermore, the target detection dataset is a dataset that has been processed by motion blur, Gaussian blur, or granular blur.

[0034] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a target detection method for a deblurred image.

[0035] As can be seen from the above technical solution, compared with the prior art, this invention discloses a target detection method for deblurred images. Using YOLOv8n as the baseline model, it introduces a High-level and Low-level Feature Block Fusion Module (HLBFM), employing an advanced Gather-and-Distribute Mechanism (GD) combined with convolutional operations and self-attention mechanisms, significantly enhancing the fusion effect of multi-scale features. Furthermore, it uses the C2f_CReToNeXt module to connect multiple sub-networks in parallel, each focusing on a specific scale of feature extraction, thereby achieving effective capture of multi-scale target information. This makes it more adaptable and exhibiting stronger detection performance when dealing with targets of different sizes and shapes in complex scenes. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0038] Figure 2 This is a schematic diagram of the blurred image detection model structure provided by the present invention;

[0039] Figure 3 A schematic diagram of the Low-GD structure for the low-level collection and distribution branch provided by this invention;

[0040] Figure 4 A schematic diagram of the information injection module structure provided by this invention;

[0041] Figure 5 A schematic diagram of the High-GD structure for the advanced collection and distribution branch provided by this invention;

[0042] Figure 6(a) is a schematic diagram of the CReToNeXt structure provided by the present invention;

[0043] Figure 6(b) is a schematic diagram of the reparameterized 3*3 inverse operation in the C2f_CReToNeXt structure provided by the present invention;

[0044] Figure 7 A comparison chart of the mAP@.5 performance of YOLOv8n and the deblurred image target detection model provided in this invention;

[0045] Figure 8 This is a flowchart of a blurred image target detection system provided in an embodiment of the present invention. Detailed Implementation

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See Figure 1 This invention discloses a target detection method for deblurred images, comprising:

[0048] Obtain the deblurred object detection dataset and the YOLOv8n baseline model;

[0049] The YOLOv8n baseline model is structurally optimized to obtain an optimized detection model;

[0050] The optimized detection model is trained using the target detection dataset to generate a deblurred image detection model;

[0051] When a deblurred image to be detected is received, the features of the deblurred image are extracted and the target location and category are predicted by the deblurred image detection model.

[0052] For details, see Figure 2 The deblurred image detection model shown in this invention, compared to the original YOLOv8n algorithm, replaces the original neck network design with a high-level feature and low-level feature block fusion module (HLBFM). This module, through a unified module architecture, effectively collects and fuses information from various feature levels, then redistributes this synthesized information to different levels, avoiding the information loss problem common in traditional feature pyramid network structures and enhancing the fusion efficiency of some features. Furthermore, a C2f_CReToNeXt module is introduced. This module performs differential processing by dividing features into two independent branches, each constrained by its own specific convolution operation. This design allows the module to capture a broader set of image features, ensuring the integrity of feature representation. After these differential operations, the outputs of the two branches are concatenated to maintain the richness and diversity of the extracted features. Combining these two modules, this invention achieves a significant improvement in the accuracy of target detection in blurred images.

[0053] Specifically, in the YOLO series of algorithms, the neck design follows the classic FPN architecture, which contains multiple branches and aims to achieve cross-scale feature fusion. However, it mainly relies on direct feature fusion between adjacent layers, while information between non-adjacent layers is obtained indirectly through recursion. This transmission mode has limitations and may lead to information loss in the computation process because inter-layer interaction is limited to information filtered by intermediate layers, and this information is not discarded during transmission. This limits the full assistance of information to specific layers, especially weakening the effective contribution to other global layers, thus potentially limiting the overall performance of information fusion. To overcome the information loss problem in the traditional feature pyramid structure during transmission, this invention constructs a high-level feature and low-level feature block fusion module, HLBFM, using a novel GD mechanism. This module, through a unified architecture, can comprehensively collect and fuse information from various layers, and then redistribute this integrated information to different layers. This not only avoids the inherent information loss problem in traditional FPN but also further enhances the ability of partial feature fusion while ensuring appropriate control of computational latency.

[0054] Specifically, the high-level and low-level feature block fusion module HLBFM includes a low-level feature collection and distribution branch and a high-level feature collection and distribution branch, which work together to achieve the above objectives.

[0055] The low-level collection and distribution branch of Low-GD mainly consists of a low-level feature alignment module (Low-FAM), a low-level information fusion module (Low-IFM), and an information injection module, such as... Figure 3 As shown.

[0056] Specifically, within the low-level feature alignment module, average pooling is used to downsample the input features to achieve uniform size. This step ensures effective information integration and significantly reduces the complexity of subsequent operations in the Transformer module. As for the low-level information fusion module, the design incorporates multi-layer re-parameterizable convolutional blocks (RepBlocks) and segmentation operations. Specifically, RepBlock takes (channels) as input and generates (channels). The intermediate channels are adjustable to adapt to diverse model specifications. The features generated by RepBlock are then segmented into two parts at the channel level, and then fused with multi-layer features, as shown in Equation 1.

[0057]

[0058] To more efficiently integrate global information into various levels, an information injection module was introduced to achieve information fusion, such as... Figure 4 As shown. Specifically, this module receives two types of input: local feature information F local With global injection information F inj For F inj The process is performed using two independent convolutional layers to obtain F. global_embed With F act As for F local F is then calculated using a single convolution. local_embed Subsequently, the fused feature representation F is calculated using an attention mechanism. out Given F local With F global Due to differences in size, average pooling or bilinear interpolation techniques are employed, based on F... inj The size of F global_embed and F act Scaling is performed to ensure precise alignment between the two. After each attention fusion step, a reparameterizable convolutional block (RepBlock) is added to further extract and fuse information. Upon entering the second stage, F... local equals B i As shown in Equation 2.

[0059]

[0060] The High-GD branch, which integrates features generated by Low-GD, has an architecture encompassing the High-FAM advanced feature alignment module, the High-IFM advanced information fusion module, and an information injection module. The specific structure is as follows: Figure 5 As shown.

[0061] The core of the advanced feature alignment module lies in average pooling, which aims to reduce the dimensionality of input features to a uniform standard. Specifically, when faced with a feature of size {R}... P3 ,R P4 ,R P5 The input features are reduced to the minimum size within each group by average pooling. Since the Transformer module is responsible for extracting high-level information, pooling operations not only facilitate the efficient aggregation of information, but also reduce the computational burden on subsequent processing steps of the Transformer module.

[0062] The advanced information fusion module integrates the Transformer architecture and feature splitting mechanism to achieve efficient modeling and fusion of global and local features. Its workflow mainly consists of three stages:

[0063] Using Transformer blocks to align the F-axis of the high-level segment feature module align Combine to generate F fuse .

[0064] F is obtained through a 1×1 convolution. fuse The number of channels was adjusted to C. P4 With C P5 The and match.

[0065] (3) F in the channel dimension fuse Segmentation is performed to obtain F inj_N4 and F inj_N5 These two parts will be fused with the features of the current level, as shown in Equation 3.

[0066]

[0067] The Transformer fusion module is built upon multiple stacked Transformer blocks, the number of which is specified by the parameter L. Each Transformer block contains a multi-head attention block, a feed-forward neural network (FFN), and residual connections. To improve inference speed, LayerNormalization, which affects speed, is replaced with Batch Normalization after convolution, and all GELU activation functions are replaced with ReLU. These changes aim to minimize the negative impact of the Transformer module on model speed. To enhance the local awareness of the Transformer blocks, a deep convolutional layer is inserted between two 1×1 convolutional layers, and the FFN's expansion coefficient is set to 2, thus achieving a balance between speed and computational cost.

[0068] The information injection module in High-GD is completely consistent with that in Low-GD. At this stage, F... local equals P i As shown in Equation 4.

[0069]

[0070] In one specific embodiment, the structure of the C2f_CReToNeXt module is shown in Figure 6(a). Its core, as a specially designed convolutional neural network component, focuses on performing complex image feature extraction tasks. The module's design philosophy is based on its unique branching mechanism. Upon receiving input, the C2f_CReToNeXt module efficiently divides the input into two independent branches, each following a specific convolution operation. This differentiated operation strategy enables the module to capture a wider range of image features, ensuring the integrity of the feature representation. After these operations are completed, the outputs of the two branches are merged, preserving the richness of the feature extraction.

[0071] Specifically, the C2f_CReToNeXt module features enhanced scalability and robustness. As shown in Figure 6(b), this is achieved by stacking multiple instances of the Rep_3×3_Reverse module, with each instance building upon the previous one to further refine the extracted features. To meet the needs of different image scales and ensure the module's adaptability, optional SPPs can be introduced to enhance its feature extraction capabilities under different image scales. During branch output merging, subsequent convolutional layers perform channel adjustment, ultimately optimizing and outputting the feature set.

[0072] The sophisticated and meticulous design of the C2f_CReToNeXt module not only enhances the operational performance of object detection but also strengthens its ability to capture and represent complex semantic information in images, making it a valuable tool in the field of object detection.

[0073] In one specific embodiment, training the model includes:

[0074] During model evaluation, multiple metrics were used, including precision (P), average precision (AP), mean average precision (mAP), recall (R), total number of parameters required by the model (Params), and Giga Floating-point Operations (GFLOPs), which measure the floating-point operations during the model's inference phase. These metrics together constitute a comprehensive assessment of model performance. The specific calculation formulas for each metric are shown in Equations 5, 6, 7, and 8.

[0075]

[0076] Among them, True Positive (TP) represents the number of samples correctly predicted as positive; False Positive (FP) refers to the number of samples incorrectly predicted as positive; and TN corresponds to the number of samples correctly predicted as negative.

[0077]

[0078] Specifically, given the scarcity of publicly available blurred object detection datasets and the limited application scenarios, this embodiment selected the Pascal VOC2007+2012 public dataset and applied three different blurring methods to the images in this dataset. After processing, a deblurring algorithm was applied to sharpen the images, and finally, an object detection task was performed on the deblurred images.

[0079] Specifically, this dataset broadly covers 20 different object categories, with each image accompanied by detailed annotations, encompassing 20 major categories including people, animals, vehicles, and furniture, ensuring that all images possess the necessary annotation information for the object detection task. This experiment utilizes the blurred Pascal VOC2007+2012 public dataset for training and validating the object detection model's performance. The training set uses the training portion of Pascal VOC2007+2012, containing 16,551 images; while the test set uses the Pascal VOC2007 test set, containing 4,952 images.

[0080] In one specific embodiment, the configuration environment is as follows: the operating system is Windows 10, the processor is an Intel(R) Core(TM) i5-10400F with a clock speed of 2.90GHz, and the GPU is an Nvidia GeForce RTX 2080Ti with 22GB of video memory. The programming language is Python 3.8, and the deep learning platform is PyTorch-GPU version 2.0.1. Furthermore, the CUDA version is 12.1, and the cuDNN version is 7.6.5.

[0081] In this experiment, the batch size for training was set to 64, and the entire training dataset was traversed 300 times (epochs set to 300). The learning rate was adjusted to 0.01. Furthermore, the input image size was uniformly set to 640 pixels, and 8 worker processes were used during data loading to accelerate data preprocessing.

[0082] To verify the effectiveness of HLBFM, this embodiment adopted the Asymptotic Feature Pyramid Network for Object Detection (AFPN), the Multiscale Separated and Enhancement Attention Module (MultiSEAM), and the Generalized Feature Pyramid Network (GFPN), and used these three networks to completely replace the original neck network. The experimental results are shown in Table 1. The data shows that although HLBFM has a higher number of parameters and floating-point operations than the other three networks, it exhibits a significant advantage in the key detection accuracy metrics. Specifically, HLBFM's mAP@.5 is 1.9% higher than AFPN, 1.5% higher than MultiSEAM, and 0.5% higher than GFPN. In the mAP@0.5-0.95 range, HLBFM shows a 1.2% improvement over AFPN, a 3.4% improvement over MultiSEAM, and a 0.2% improvement over GFPN. Based on the combined data, it can be clearly concluded that HLBFM can more accurately identify targets when detecting blurred images, thus better meeting the actual needs of high-precision detection.

[0083] Table 1 Comparison of neck network improvements

[0084]

[0085] Because blurred images are often accompanied by high noise and unclear edge boundaries, this experiment completely replaced the C2f module in the backbone network and conducted in-depth comparative research. Detailed experimental results are shown in Table 2. Through detailed comparison of each module, it was found that although C2f_EFAttention, C2f_CAA, C2f_MuliDilatelocalAttention, C2f_LSKAAttention, and C2f_MLCA ​​maintain a lower level in terms of parameter scale and floating-point computation compared to C2f_CReToNeXt, they show a significant gap in detection accuracy. Specifically, C2f_CReToNeXt exhibits better performance in handling complex features of blurred images, with mAP@.5 reaching 42.9% and mAP@0.5-0.95 reaching 29.2%. This data strongly verifies that the C2f_CReToNeXt module has good complex feature extraction capabilities when dealing with blurred images.

[0086] Table 2 Comparison of Improvements to the Backbone Network C2f Module

[0087] Module mAP@.5 / % mAP@0.5-0.95 / % Params / M FLOPs / G Baseline 41.4 28.1 3.0 8.2 C2f_EFAttention 39.2 25.4 2.6 6.9 C2f_CAA 40.5 26.9 2.6 7.0 C2f_MuliDilatelocalAttention 40.2 26.4 2.6 7.1 C2f_MLCA 42.7 29.0 3.0 8.1 C2f_LSKAAttention 40.9 27.0 2.6 7.4 C2f_CReToNeXt 42.9 29.2 4.1 11.0

[0088] In one specific embodiment, to further verify the effectiveness of the HLBFM and C2f_CReToNeXt modules of the present invention, an ablation experiment was designed, and the experimental results are summarized in Table 3. In this experiment, the YOLOv8n model was used as the baseline model. The experimental data shows that when the original neck network structure was replaced by HLBFM, mAP@.5 and mAP@0.5-0.95 reached 43.4% and 29.3%, respectively. Furthermore, when the C2f module in the backbone network was replaced by the C2f_CReToNeXt module, mAP@.5 and mAP@0.5-0.95 achieved improvements of 1.5% and 1.1%, respectively. While employing these two innovations simultaneously, although the number of parameters and floating-point operations increased compared to the baseline model, the improved detection algorithm achieved a significant improvement in overall detection accuracy, with mAP@.5 and mAP@0.5-0.95 increasing by 2.3% and 1.6%, respectively. Therefore, it is evident that the detection algorithm proposed in this invention effectively helps to improve the detection accuracy of blurred images.

[0089] Table 3 Comparison of ablation experiments

[0090] HLBFM C2f_CReToNeXt mAP@.5 / % mAP@0.5-0.95 / % Params / M FLOPs / G 41.4 28.1 3.0 8.2 √ 43.3 29.3 6.0 12.1 √ 42.9 29.2 4.1 11.0 √ √ 43.7 29.7 7.1 15.3

[0091] Specifically, under the same experimental conditions, the performance comparison results of the object detection method for deblurred images proposed in this invention with other mainstream object detection algorithms are shown in Table 4. Compared with the two-stage object detection algorithm, although the mAP@0.5-0.95 of this invention is 2.1% lower than Cascade R-CNN, its mAP@.5 is 2.3% and 2.9% higher than Cascade R-CNN and Faster R-CNN, respectively. Compared with the single-stage object detection algorithm SSD, the object detection method for deblurred images proposed in this invention improves mAP@.5 and mAP@0.5-0.95 by 4.1% and 2.1%, respectively; although its detection accuracy is slightly lower than RT-DETR, the object detection method for deblurred images proposed in this invention has a much lower number of parameters and floating-point operations than RT-DETR. Compared with the YOLO series algorithms, the object detection method for deblurred images proposed in this invention has better accuracy than most YOLO series algorithms, and its mAP@.5 is 2.1% lower than that of the YOLOv8s large model. Compared to the baseline model YOLOv8n, the accuracy of this invention is significantly improved, with mAP@.5 and mAP@0.5-0.95 improved by 2.3% and 1.6%, respectively. While significantly improving accuracy, the number of parameters and floating-point operations of the algorithm only increases slightly, thus ensuring that the algorithm maintains good performance in blurred image detection scenarios.

[0092] Table 4 Performance Comparison of Different Target Detection Algorithms

[0093] Model mAP@.5 / % mAP@0.5-0.95 / % Params / M FLOPs / G CascadeR-CNN 41.4 31.8 69.1 236 FasterR-CNN 40.3 28.5 41.3 208 SSD 39.6 27.6 24.0 30.5 RT-DETR 44.4 30.4 32.8 110.2 YOLOv8n (baseline) 41.4 28.1 3.0 8.2 YOLOv5n 41.2 27.3 2.5 7.1 YOLOv5s 41.6 29.5 9.1 24.1 YOLOv6n 41.3 28.1 4.2 11.9 YOLOv6s 41.7 30.3 16.3 44.2 YOLOv7tiny 43.5 26.3 6.0 13.2 YOLOv8s 45.8 31.9 11.1 28.5 YOLOv9t 43.2 29.6 2.0 7.9 YOLOv10n 42.8 29.2 2.7 8.4 This article 43.7 29.7 7.1 15.3

[0094] In one specific embodiment, to more intuitively demonstrate the improvement in detection performance between the present invention and the YOLOv8n baseline algorithm, the resulting images of the two algorithms were compared in a paired manner. When facing blurry image detection tasks, the YOLOv8n algorithm exhibited multiple missed detections, while the present invention successfully captured these targets missed by YOLOv8n. This phenomenon strongly proves that, compared to YOLOv8n, the present invention has a more powerful target detection capability and can accurately identify objects that YOLOv8n struggles to detect. Furthermore, in terms of the core indicator of detection accuracy, the present invention also demonstrates a significant advantage, with its overall performance generally surpassing that of the YOLOv8n algorithm.

[0095] like Figure 7 As shown, compared with the YOLOv8n model, the deblurred image target detection model of the present invention shows a continuous and significant increase in detection accuracy, which fully verifies its excellent detection capability.

[0096] Specifically, this invention addresses the problems of complex noise, blurred boundaries, missing feature details, and insufficient detection accuracy faced by blurred images, proposing a blurred image target detection algorithm based on an improved YOLOv8n. In this algorithm framework, HLBFM is incorporated into the backbone network to reduce information loss during computation and enhance feature fusion capabilities. Simultaneously, the neck network replaces the original C2f module with the C2f_CReToNeXt module to further improve feature extraction accuracy. Under the same experimental conditions, comparative experiments with the original algorithm clearly demonstrate that the improved algorithm performs better in blurred image detection tasks, with a significant improvement in detection accuracy.

[0097] On the other hand, the present invention discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a target detection method for a deblurred image.

[0098] On the other hand, see Figure 8 This invention also discloses a blurred image target detection system. The main functional requirements of the system are as follows: First, the system needs to provide user registration and login functions. After logging in, the system should support both video analysis and image analysis modes, allowing users to upload a single blurred image as needed. For the uploaded blurred image, the system needs to call an image deblurring model to perform deblurring processing, restoring the image's clear details and providing high-quality input for subsequent target detection. The deblurred image needs to be further fed into a deblurred image target detection model for target detection, identifying target objects in the image and labeling their location, category, and confidence level.

[0099] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A target detection method for deblurred images, characterized in that, include: Obtain the deblurred object detection dataset and the YOLOv8n baseline model; The YOLOv8n baseline model is structurally optimized to obtain an optimized detection model; The optimized detection model is trained using the target detection dataset to generate a deblurred image detection model; When a deblurred image to be detected is received, the features of the deblurred image are extracted and the target location and category are predicted by the deblurred image detection model.

2. The target detection method for deblurred images according to claim 1, characterized in that, Structural optimization of the YOLOv8n baseline model includes: The neck network of the YOLOv8n baseline model is replaced with a high-level feature and low-level feature block fusion module HLBFM, which is used to collect and fuse feature information from all levels and redistribute it to different levels. The C2f module in the backbone network of the YOLOv8n baseline model is replaced with the C2f_CReToNeXt module, which is used to extract multi-scale features by connecting multiple sub-networks in parallel.

3. The target detection method for deblurred images according to claim 2, characterized in that, The high-level features and low-level features are divided into blocks. HLBFM Fusion Module include: The Low-GD branch, which collects and distributes low-level features, is used for downsampling and feature fusion of input features. The High-GD branch, which focuses on advanced feature collection and distribution, is used for global information extraction and fusion of input features.

4. The target detection method for deblurred images according to claim 3, characterized in that, The Low-GD branch for low-level feature collection and distribution includes: Low-FAM (low-level feature alignment module), Low-IFM (low-level information fusion module), and a first information injection module; The Low-FAM low-level feature alignment module downsamples the input features using average pooling. The Low-IFM low-level information fusion module achieves feature fusion through reparameterizable convolutional blocks (RepBlock) and segmentation operations. The first information injection module is used to perform attention fusion between local feature information and global injected information.

5. The target detection method for deblurred images according to claim 4, characterized in that, The High-GD branch of advanced feature collection and distribution includes: High-FAM (High-Face-Am), High-IFM (High-IFM), and a second information injection module; The High-FAM module reduces the input features to a uniform size through average pooling. The High-IFM (High-Information Fusion) module extracts and fuses global features through the Transformer block. The second information injection module is used to perform attention fusion between local feature information and global injected information.

6. The target detection method for deblurred images according to claim 5, characterized in that, Both the first information injection module and the second information injection module include: The globally injected information is processed through two independent convolutional layers to obtain the first feature and the second feature; The local feature information is processed through a convolutional layer to obtain the third feature; The first feature, the second feature, and the third feature are fused using an attention mechanism.

7. The target detection method for deblurred images according to claim 2, characterized in that, The C2f_CReToNeXt module includes: The branching processing mechanism divides the input features into two independent branches, and performs differential convolution operations on each branch. The feature merging module is used to merge the outputs of two branches.

8. The target detection method for deblurred images according to claim 1, characterized in that, The target detection dataset is a dataset that has been processed by motion blur, Gaussian blur, or granular blur.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a target detection method for a deblurred image as described in any one of claims 1 to 8.

Images