Research And Implementation Of Target Detection Based On YOLOv3 Algorithm

Due to the explosive growth of computer performance and data,deep learning algorithms came into being.There are many fields of deep learning algorithm research,including natural language processing,autonomous driving,computer vision and so on.Target detection is one of the hot research directions in computer vision.Target detection has important research significance in security monitoring,automatic driving,face recognition,and medical imaging.With the development of deep learning,target detection algorithms based on deep learning have sprung up and proposed.They are all committed to improving the performance of the target detection algorithm and increasing the value of mAP and FPS to meet the needs of the industry.In order to improve the detection accuracy of the target detection algorithm,this paper is based on IOU(a new loss function L_IOU is proposed.In order to further optimize the L_IOU loss function,this paper first proposes IIOU(based on IIOU and a new loss function L_IIOU.With the new proposal The L_IOU loss function and the L_IIOU loss function replace the original loss function in the YOLOv3 algorithm.Tests on the MS COCO data set show that the two newly proposed loss functions both improve the average detection accuracy of the YOLOv3 algorithm.To reduce the phenomenon of missed inspections.The loss functions of the above mentioned target detection algorithms are all L1 or L2 loss functions.So this paper proposes the L_IOU loss function to replace the original loss function in the YOLOv3 algorithm.Experiments on the MS COCO data set show that the mAP value is 59.38,the original YOLOv3 algorithm mAP value is 57.9,and L_IOU as the loss function of YOLOv3 is 1.48 higher than the original YOLOv3 algorithm mAP value.But IOU still has problems.Then this paper proposes IIOU to solve the problems of IOU,and then proposes a new loss function L_IIOU to optimize the L_IOU loss function.Replace the loss function in the YOLOv3 algorithm with L_IIOU,and perform an experiment on the MS COCO data set.The mAP value is 60.77,which is 1.39 higher than L_IOU as the loss function of the YOLOv3 algorithm.The mAP value is 2.87 higher than the original YOLOv3 algorithm.In order to improve the robustness of the system,the YOLOv3 algorithm was trained on the augmented data set.This article finally implements a target detection system based on the improved YOLOv3 algorithm,which enables users to upload videos or pictures,complete the target detection annotation,and return the results to the user.At the same time,the quality of the system engineering is strictly controlled.After testing,the system functions are in a stable state.