Image Object Detection Method Based On R-CNN

Object detection is an important research in computer vision and pattern recognition.Many scholars have conducted in-depth research in this direction and have achieved fruitful results through continuous improvement and optimization.With the development of deep learning,at present,Object detection algorithms have changed from traditional algorithms to deep learning algorithms,including two-stage Object detection algorithms and one-stage Object detection algorithms.In view of the fact that the Faster R-CNN algorithm in the two-stage target detection has the advantages of high detection accuracy and model stability,but detection of targets of different sizes and different poses still has problems such as misdetection and missed detection.This thesis focuses on the Faster R-CNN algorithm.The specific contents are as follows:(1)In view of the fact that the appearance of the Object in the image will vary greatly according to its basic shape,different postures and perspectives,this thesis proposes a multi-channel detection algorithm.Firstly,the original single detection channel is changed to a network composed of three channels for training and testing,which effectively improves the accuracy of Object detection;Second,use the minimum regularization function and loss function to optimize the network,which improves the detection performance and saves memory space;Finally,in terms of experimental verification,this article not only conducted experimental verification on multiple mainstream data sets,but also tested and verified some daily scene images shot on the spot.The experimental results show that: in the PASCALVOC2012 data set,the average accuracy of the algorithm in this paper is 79.1% when the training time is slightly increased,which is improved to varying degrees compared with the YOLOv3 and Faster R-CNN algorithms.(2)Taking into account the huge differences in Object size,shape,occlusion and lighting conditions in Object detection tasks,the Faster R-CNN algorithm is improved.First of all,this article uses ResNet-101 network to replace the original VGG network for feature extraction,which can further improve the Object detection performance;Secondly,the RoI Align unit is introduced and combined with the bilinear interpolation method to retain floating-point numbers to improve detection accuracy;Finally,the soft-NMS algorithm is used to replace the traditional NMS algorithm for non-maximum suppression,so that objects with high overlap can be successfully detected.The experimental results show that the average accuracy of the improved algorithm in the PASCALVOC2007 data set is 80.2%,which is 3.8% higher than the faster R-CNN algorithm.