Research On Object Detection Of Haze Images Based On Deep Learning

As one of the hot research directions in the field of computer vision,object detection plays an important role in the fields of robot navigation and unmanned driving.At present,various object detection algorithms have achieved good results on public datasets.However,haze weather in practical application scenarios will affect the quality of images collected by electronic devices,resulting in increased noise in images,grayish white images,and decreased contrast.These factors seriously affect the performance of the object detection model,resulting in a decrease in detection accuracy and then affecting the application of subsequent vision tasks.Therefore,in view of the problem of low object detection accuracy in haze weather,based on the YOLOv5 object detection algorithm,the following research work is carried out:(1)Aiming at the situation that the characteristics of objects in haze images are difficult to identify and the poor visual effect of pictures,the Retinex series of image enhancement algorithms are studied in this paper.And a large number of haze images are used to compare and analyze the dehazing effects of two classic image enhancement algorithms,AMSRCR and MSRCP.Experiments show that the images processed by the MSRCP algorithm show more balanced enhancement and better visual effects in many aspects such as peak signal-to-noise ratio and structural similarity.The MSRCP algorithm can effectively reduce the influence of haze and improve the clarity of the image.(2)Aiming at the problems of wrong detection and missed detection in the object detection task in haze weather,in order to reduce the negative impact of haze on object detection as much as possible,a strategy of dehazing algorithm + object detection algorithm is proposed in this paper.The YOLOv5 n object detection algorithm is selected to conduct comparative experiments on the haze image dataset and the dataset dehazed by the MSRCP image enhancement algorithm.The experimental results show that the mAP of YOLOv5 n model to detect the dehaze image after MSRCP image enhancement is 3.2% higher than that of the original haze image,which verifies the effectiveness of the object detection YOLO algorithm + dehaze algorithm strategy proposed in this paper.(3)In order to reduce the calculation amount of the model and improve the real-time detection ability of the model,an improved object detection algorithm YOLOv5s-F is proposed in this paper,which adopts a strategy to add the Ghost Bottleneck module of the SE attention mechanism to optimize the Back Bone of YOLOv5 s.For the situation that the aspect ratio of the Bounding Box is not considered in the regression process and affects the accuracy,CIoU is selected as the regression loss function of the model.In order to optimize the Box＿loss loss,the K-means algorithm is chosen to optimize the Anchor box,which significantly improves the detection ability and convergence speed of the model.For the object detection task in the low-visibility environment affected by haze,a research of object detection on a dataset with 20 categories is carried out in this paper.The experimental results show that the mAP of the strategy of MSRCP dehazing algorithm +YOLOv5s-F object detection model proposed in this paper is 2.9% higher than that of using the original YOLOv5 s model to directly detect low-quality images with haze,and the GFLOPs is reduced by 67.5%,the inference speed has increased by 45.8%.The strategy in this paper significantly improves the real-time detection capability of the YOLOv5 model,and has strong practical application significance.