Research On Object Detection With Deep Learning

Deep learning,as an important branch of artificial intelligence,is now attracting more and more attention,and target detection,a basic research field of computer vision,has also shown new vitality due to the rapid development of deep learning.Target detection is to find the position of the target object in the image or video and determine the category of the object.Due to the different shapes,sizes,numbers and positions of targets in images,target detection has always been one of the difficult problems in the field of computer vision detection.Traditional target detection adopts sliding window and image scaling,which has poor detection efficiency and low accuracy.In order to improve the disadvantages of traditional detection algorithm,a method combining depth learning and target detection is applied.There are two kinds of target detection algorithms based on depth learning.One is twostage target detection algorithm represented by R-CNN series.This kind of algorithm first generates candidate regions,then classifies the candidate regions and corrects the border positions.The other is one-stage target detection algorithm represented by SSD,YOLO,etc.This kind of algorithm does not need to generate candidate regions to directly regression the target objects.In this paper,Faster R-CNN and SSD are two kinds of target detection algorithms that have been deeply studied.The specific contents are as follows:(1)Deep learning technology and target detection algorithm are studied,and the working principles of convolution neural network and commonly used target detection technology are analyzed.(2)The algorithm principle of Faster R-CNN is deeply studied through the network structure,anchor frame generation and training process.Aiming at some problems existing in the original algorithm,this paper adds a feature pyramid network to the original network structure and optimizes the classification labels of target categories.In addition,the experimental training images are transformed differently to expand the data set,which makes the detection effect of the original algorithm optimized.(3)The SSD algorithm is studied,and the overall structure of SSD,the generation of prior frames,the matching of prior frames,the mining of hard samples and the loss function are analyzed in depth.In order to improve the original algorithm,the Top-Down module is added to the network of the original algorithm,the classification loss is optimized,and the Soft-NMS algorithm is used.(4)The deep learning framework MXNet is used on the computer platform to carry out training experiments before and after improvement on Faster R-CNN algorithm and SSD algorithm,and the test results of the experiments are analyzed.The results show that the test results after improvement have been improved to a certain extent compared with those before improvement.