| In recent years,object detection has been widely applied in various fields such as face recognition,autonomous driving and intelligent surveillance.The core task of object detection is to determine the location and category of the targets.The quality of the detection image and the network model structure are important factors that affect the accuracy of object detection algorithms,whether it is the two-stage object detection algorithm represented by Faster R-CNN or the one-stage object detection algorithm represented by the YOLO series.In foggy weather conditions,the lower contrast of the image and the loss of target details make it easy to miss or misidentify targets when using existing object detection algorithms directly.Based on the above background,this thesis conducts research on object detection algorithms under foggy-weather conditions,and the main work is as follows:(1)To address the problems of insufficient detection accuracy and low-quality detection boxes in traditional object detection algorithms,this thesis proposes an improved object detection algorithm based on YOLOv5 s,named YOLOv5s-Enhanced.Firstly,a coordinate attention mechanism is introduced to enhance the accuracy of target localization by encoding channel information and spatial information.Secondly,the traditional Bottleneck structure is improved by designing the fine-grained residual block CA_Res2Block with the fusion attention mechanism to fuse local multi-scale features in a fine-grained manner,increase the internal receptive field of the model,and improve the neural network’s feature extraction ability.Then,the Neck structure of the model is improved by designing the Cross-Layer Weighted Feature Fusion module(CWFF)to enhance the network’s focus on shallow detail features.Finally,the original loss function is replaced with the CIo U loss function,and a k-means++ clustering method and multi-scale training mechanism are combined to further enhance the model’s robustness and improve its detection performance.Experimental results show that the proposed method achieves a m AP value of 88.15% on the PASCAL VOC dataset,which is 2.16% higher than that of YOLOv5 s.(2)This thesis introduces an object detection algorithm in foggy weather called JOA-ODFC(Joint Optimization Algorithm for Object Detection under Foggy-weather Conditions)that integrates a dehazing module to address the issue of low image quality under foggy-weather conditions leading to decreased detection accuracy.By jointly optimizing the algorithm,the dehazing module’s core task is shifted from visual dehazing to feature map generation to better serve subsequent detection tasks.Firstly,the Improved-AOD-Net is designed as a deep dehazing network based on AOD-Net,which incorporates a hybrid dilated convolution module and a feature fusion module.Secondly,based on YOLOv5s-Enhanced and Improved-AOD-Net,JOA-ODFC is designed as a joint optimization foggy weather object detection algorithm.Finally,to address the issue of limited sample size in foggy weather datasets,this thesis designs a foggy image generation algorithm based on a depth estimation model,guided filtering,and atmospheric scattering models to expand the dataset using cleaned PASCAL VOC and Cityscapes data.The results of several experiments demonstrate that JOA-ODFC outperforms other foggy weather detection algorithms in terms of detection accuracy and effectiveness.(3)This thesis designs and implements a foggy weather object detection system based on JOA-ODFC.The system requirements,system architecture,and system hardware parameters are introduced,and the detection results of foggy images are demonstrated on both the web and mobile platforms. |
