Research On Object Detection Algorithm Based On YOLOV4

Object detection is a key technology in the field of computer vision,which aims to find objects of interest from images,and determine the category and location of objects in images.Object detection has a wide range of applications in intelligent monitoring,bioinformatics,automatic driving and other fields,and is also a hotspot of academic research.How to further improve the accuracy and efficiency of object detection is an urgent problem to be solved.Aiming at this problem,this thesis studies the data processing and object detection algorithm respectively,and proposes a new data processing algorithm and an improved YOLOV4 algorithm.The innovations of the this thesis are as follows:(1)The data set contains out-of-distribution data,but existing object detection algorithms may misjudge the out-of-distribution data as in-distribution data with high confidence during detection,resulting in incorrect detection results and reducing the effectiveness of the detection algorithm.Therefore,from the perspective of data processing,a new out-of-distribution data detection algorithm,Noise Teaching for OOD Detection(NTOOD),is proposed.The algorithm regards out-of-distribution data in the training set as noise data,and uses Co-teaching for training.And the method of removing the noise proportion in a proportional decreasing manner is used to eliminate the influence of noise data on the algorithm,so as to achieve the purpose of improving the detection accuracy.The experimental results on multiple image datasets such as Gaussian Noise,Blob,Texture,and SVHN show that the NTOOD algorithm can effectively identify out-of-distribution data,and its detection accuracy is better than that of Baseline,Deep-MCDD and other out-of-distribution data detection algorithms and Co-teaching algorithm.(2)The YOLOV4 algorithm based on convolutional neural network combines multi-scale feature fusion and YOLO detector,and has good object detection performance.In order to further improve the object detection accuracy of the algorithm,an improved YOLOV4 algorithm,YOLOV4TB(YOLOV4+Transformer+Bi FPN),is proposed.The algorithm uses Transformer to extract the features of objects that occur together on the same feature map for feature enhancement to improve the detection accuracy of small target objects and blocked objects;in the feature fusion stage,the Bi FPN(Bi-directional Feature Pyramid Network)module is used to solve the problems of redundant calculations and failure to consider the different contribution of different feature layers in the PAN module of YOLOV4.On this basis,the Leaky-Relu activation function and depth separable convolution technology are adopted to solve the problem of the decrease in object detection accuracy and the increase in the amount of parameters and computation.question.The experimental results on the PASCAL VOC dataset show that compared with YOLOV4,the detection accuracy of the YOLOV4 TB algorithm is improved by 3.7%,especially for small objects and blocked objects.At the same time,the detection efficiency has also been improved.