Research On Target Recognition Algorithm Based On Convolutional Neural Network

Object detection is one of the main research directions of current computer vision.Its main realization method is to realize object category judgment and object position recognition through images or video.At the same time,object detection technology is widely used in medical,military and transportation fields application.In recent years,the technology based on deep learning has penetrated into all walks of life,especially the convolution neural network technology developed under the background of deep learning has shown unprecedented model advantages in the field of image processing.Compared with the traditional target detection algorithm,the target detection algorithm based on convolution neural network has the advantages of faster speed and higher detection accuracy,so it has become the mainstream algorithm in the current target detection field.Single-stage detectors such as YOLO have a faster detection speed than two-stage detectors such as Faster-RCNN,so they are more suitable for real-time detection tasks,Although single-stage detectors such as YOLO are object detection algorithms based on regression models,they have the problem of low detection accuracy.For the problem of low accuracy of single-stage detectors,thesis focuses on sample balance and borders Regression and other aspects are researched,analyzed and improved based on the YOLOv3 algorithm.First of all,this article introduces the theoretical basis of the object detection algorithm based on convolutional neural networks,and describes in detail the processing process of convolutional neural networks on images and the network structure of convolutional neural networks.Secondly,the detection accuracy of the YOLOv3 algorithm is not high.Because the YOLOv3 algorithm is a detection algorithm based on regression models,in other words,the algorithm extracts and learns features by resampling the input data,because most of information of an image belong to the background information,and the target object information that needs to be learned is too small compared to the background information.As a result,the parameter update of the model during training is mainly dominated by the background sample and the simple sample.Thesis uses a sample balance method to relieve the negative impacts of effective information being suppressed due to unbalanced distribution.By further reducing the loss coefficients of negative samples and simple samples in the loss function to suppress the negative effects of negative samples and simple sample losses in the training process,while increasing positive samples and difficult samples to increase the importance of both parameters in the training process.Compared with the YOLOv3 algorithm on the Pascal VOC test dataset,the improved YOLOv3 algorithm reduces the classification error of the algorithm and improves the detection accuracy of the algorithm.Finally,the accuracy of the bounding box regression in the target detection algorithm has a great impact on the performance of the final detection algorithm.For the GIOU(Generalized Intersection over Union)bounding box regression,it degenerates into an IOU loss when the bounding box is in the containment relationship and it needs more time to convergence when the bounding box is in a parallel and vertical position relationship in space.In order to solve these problems,thesis uses CIOU(Complete IOU)border regression for reference on the basis of GIOU,and integrates the spatial distance of the center point of the border and border similarity into the border regression to improve the accuracy of the algorithm's bounding box regression.In the final algorithm model,the sample balance optimizer is added to the model,so that the algorithm can be optimized in both classification and positioning tasks.The experimental results show that the improved YOLOv3 algorithm has improved the convergence speed,and the recall rate and detection accuracy have been improved.