A Research Of Target Detection And Recognition Algorithm With Region-based Convolutional Neural Network

With the application of deep learning to the field of computer vision,image processing methods have achieved breakthrough development,and target detection and recognition as one of its basic tasks are currently widely used in video surveillance,face recognition,and target tracking.Therefore,improving the accuracy of target detection and recognition has also become a hot research topic.This article focuses on the object detection and recognition algorithms based on deep learning,and makes in-depth research on its principles and applications.In the deep learning-based object detection and recognition algorithms,they are roughly divided into two categories: one is based on regional convolutional neural networks;the other is based on regression.This article mainly studies the target detection and recognition algorithms based on regional convolutional neural networks(R-CNN series).This type of algorithm first generates candidate regions,then uses convolutional neural networks to extract features,and finally performs detection frames through classifiers and regressors.Among them,Faster R-CNN is an algorithm proposed by R-CNN after several improvements.It is one of the most accurate target detection and recognition algorithms,but there is still a lot of room for improvement.This article tests these series of algorithms,and in-depth studies the principle and algorithm framework of Faster R-CNN algorithm,so as to analyze its advantages and disadvantages and make improvements.This article mainly makes two improvements to Faster R-CNN algorithm.The first is to improve the Non-Maximum Suppression algorithm: Non-Maximum Suppression is a method for screening detection frames that is used by almost all target detection algorithms.Traditional Non-Maximum Suppression algorithms detect the highest classification score.The frame is used as the reference frame,and then the frame with an IoU greater than a certain value is deleted.This algorithm can effectively remove overlapping redundant frames,but it has the disadvantage that it cannot handle the situation where the two frames are adjacent.Containing two adjacent or highly overlapping target objects,the use of traditional Non-Maximum Suppression processing can easily cause the wrong detection frame to be deleted and cause missed detection.This article aims to reduce missed detections by improving the traditional Non-Maximum Suppression algorithm.The improved Non-Maximum Suppression does not directly delete the detection frame with high overlap with the reference frame,but reduces its score one by one In this way,the improved Non-Maximum Suppression is no longer as violent as the original algorithm,and the box selection is more reasonable.The second point is to propose a new multi-task loss function.The task of target detection and recognition is divided into classification and localization,so the loss function of Faster R-CNN is divided into classification loss and regression loss.There is a hyperparameter that makes this The weights of the two losses are roughly the same.This article finds that the sensitivity of large-scale targets and small-scale targets to classification and regression is not the same.Based on this,we propose a multi-parameter loss function.Its purpose is to make regression loss account for a larger weight when detecting large targets;more attention is paid to classification losses when detecting small targets.This improves the accuracy of detecting small targets and the overall accuracy of the algorithm.This article conducts experiments on public data sets and compares them with current mainstream target detection and recognition algorithms to verify the effectiveness of the algorithms.The experimental results show that the improvement of the Non-Maximum Suppression algorithm and loss function in this article improves the accuracy rate compared with the original algorithm,and also has advantages compared with mainstream target detection and recognition algorithms.