Research And Realization Of Object Detection Based On Convolution Neural Network

Object detection is a research hotspot in computer vision.Convolutional neural networks are widely used in object detection due to their rich feature learning levels and diverse feature expression capabilities.The main purpose of object detection is to detect,locate and mark targets from pictures or videos.It comprehensively uses image processing,target positioning,target recognition and other technologies,and is widely used in many fields.The existing object detection algorithms are mainly divided into two categories,one is the object detection method represented by R-CNN.The characteristic of this type of detection method is that the detection accuracy is high,but the detection speed is slow;the other type is the object detection method represented by YOLO and SSD.The detection speed is faster,but the detection accuracy is not high enough.Therefore,the current two types of object detection methods can not meet the requirements of fast detection speed and high detection accuracy.Based on the calculation complexity,detection speed and detection accuracy,based on the existing YOLO object detection algorithm,this paper studies and improves the object detection algorithm.The paper first studies the basic framework and method principles of mainstream object detection represented by R-CNN,YOLO,etc.in current object detection,analyzes some of the algorithms in the object detection process,and deeply studies the multiple factors that affect the detection speed and accuracy Various object detection methods improve the object detection algorithm based on YOLO.The commonly used Io U algorithm has been improved,and the area and distance parameters have been increased to make up for the shortcomings of losing the ability to measure the pros and cons of the prediction box in the early and late stages of training;the pyramid feature extraction network has been improved to increase the difference between different scales.The information exchange of feature maps.By changing the feature fusion method,the information of targets of different sizes in the feature map is enriched to a certain extent.This paper uses the PASCAL VOC data set as the training and verification data set.Through experimental verification,the improvement of the Io U algorithm has improved its detection effect by 4.3%,and the improvement of the feature extraction network has improved the detection effect by 3.1%.After the two aspects of optimization and improvement were integrated,the overall detection effect increased by 4.9%.Under the premise that the total complexity of the algorithm is basically unchanged,the detection performance of the object detection algorithm is generally improved.