Object Detection and Recognition based on Improved Faster R-CNN

Abstract

With the rapid development of deep learning methods, object detection, as one of the most basic and challenging tasks in the field of computer vision, has made remarkable progress. Graph convolutional network (GCN) [1] extracts image spatial features, which can effectively integrate semantic information and geometric information in a graph structure for learning. In this paper, the feature extraction network of the Faster R-CNN [2] object detection algorithm is changed from the original Vgg16 network to GCN, which improves the detection effect of the algorithm on multiple targets and improves the detection speed. In the fully connected layer of the network, GCN is used to process the target frame area extracted by the Faster-RCNN model to obtain a target frame that is more in line with the semantic information of the image, and then the final task classification and target recognition are performed to effectively utilize the essential features of the object. and intrinsic connections improve the generalization ability and robustness of the model. In the experimental comparison between PASCAL VOC 2007 [3] and Microsoft COCO [4] datasets, the detection speed of the method in this paper is increased by 12% and 13.5%, and the average detection accuracy is increased by 3.8% and 5%, which reduces the amount of calculation and improves the recognition accuracy.