Research On Object Detection Algorithm Based On Deep Learning

In recent years,with the continuous improvement of computer computing power,deep learning technology based on neural networks has gradually made breakthroughs in Computer Vision,Speech Recognition,Natural Language Processing,data mining and other fields.The current object detection algorithms based on deep learning are superior to traditional machine vision algorithms in terms of detection speed and accuracy.But for the situation that the scale of the detected object is too small,the object detection algorithm based on deep learning still has the shortcomings of high missed detection rate and low detection accuracy.In order to improve the accuracy of small object detection and reduce the missed detection rate,this paper proposes an improved version of the object detection algorithm based on YOLOv4: SO-YOLO algorithm.This algorithm can not only meet the speed and accuracy of small target object detection,but also can successfully detect manufacturing defects on the micron scale in the case of small samples.The main research work and results of this article are as follows:(1)SO-YOLO algorithm proposes a new feature fusion method,which is better than YOLOv4 in object target detection.Concretely,the low-level features extracted from the shallow part of the network contain rich location information.For small object detection,fusion of these location information becomes crucial.The small object feature fusion method designed in this research will fully integrate the high-level and low-level information of the network.In this way,the feature map obtained by fusion contains rich position information and semantic information at the same time,so that small target defects can be detected more accurately.(2)The SO-YOLO algorithm optimizes the YOLO Head part and reduces the complexity of the network model.Concretely,in the data preprocessing stage,the K-means++ clustering algorithm is used for a priori box enhancement.Then use the enhanced anchor box to predict on the feature map of the appropriate scale.Finally,delete other network branches in the YOLO Head section.(3)For small sample problems,use data augmentation to expand the data set.In the data preprocessing stage,the data set is expanded to 8 times the size of the original data set through the traditional data enhancement method,and then the mosaic data enhancement method is used to further enhance the data set.The data set used in this study is a CMOS holders produced by an industrial production line.The objects to be detected in this data set are all small targets.Compared with the YOLOv4 algorithm,the SO-YOLO algorithm proposed in this paper reduces the network model parameters by 29.29% while increasing the accuracy by 1% and the recall by 1%.