Research On Deep Learning Object Detection Method Based On Improved RPN

With the wide application of artificial intelligence technology in the field of computer vision,object detection is a frontier direction that has attracted much attention in the field of computer vision.It involves detection and recognition tasks in images and is used in medical image processing,image retrieval,automatic driving,and human Face recognition and other scenes.From the early method of combining traditional image features with machine learning,it gradually developed into a method based on deep learning.This paper researches on a twostage object detection model.In this type of model,the region proposal network(RPN)proposes positive candidates in the first stage.The samples for training at this stage are selected after matching the object position in the image with the anchor box,resulting in the model being driven by most objects with good matching conditions,resulting in a large number of missed detections.For small objects and long Objects with larger widths are more obvious;secondly,the problem of imbalanced samples in object detection tasks has a greater impact on training.To solve these problems,based on the classic model Faster R-CNN algorithm in the twostage object detection framework,this paper conducts research on RPN sample matching and training loss optimization.At the same time,the disease image data in the project is used as an example for implementation.The improvement of object detection effect brought by the strategy.The main innovations are as follows:(1)Aiming at the imbalance of sample extraction and matching in the RPN stage,an improved B-RPN(Balanced region proposal network)algorithm is proposed.The algorithm sets an adaptive loose threshold for positive samples corresponding to difficult matching objects,and sets a threshold strategy to increase the distance between positive and negative samples,improves the quality of RPN sample extraction,and improves detection performance in difficult matching classes obvious.(2)Aiming at the problem of insufficient geometric transformation modeling ability of the network and the imbalance of different types of data,a new object detection network is constructed.The network uses Res Next and feature pyramid network as backbone,which enhances the feature learning ability of the network.In order to improve the geometric transformation modeling ability of the network,it is proposed to replace part of the convolution with Deformable Convolutional Networks(DCN);for the problem of data imbalance of different types of objects,an improved B-Focal-loss loss function is proposed.Compared with traditional Focal-Loss,the loss function pays more attention to the difficulty of classifying object classes at the current stage,and drives the model to adaptively adjust the training proportion of different classes.The improved network's missed detection rate and the detection effect on small amounts of data have improved significantly.By further combining the techniques of Soft-NMS and multi-scale model fusion in the testing stage,the improved object detection network has achieved a significant improvement in accuracy.