An object classification and detection method with faster R-CNN based on multiscale anchor box and weighted cross entropy classification

Abstract

To improve the accuracy of object classification and recognition, one of the methods is to add recommendation boxes, recommendation boxes are generated by region proposal algorithms. Because of the inherent ratio, the size of the candidate box generated by Faster RCNN's RPN is usually large, which would easily cause a great number of overflows in sliding search. This is unfriendly for multi-objects detection. To improve the precision of multi-objects detection, we introduce a multiscale proposal box that could predict object bounds and object scores at each position. Then, in order to increase the positive sample range of the foreground, the weighted cross entropy classification function is used for binary classification in the RPN network. In the ROI network, the candidate frame reset algorithm is proposed to realize the position regression of the prediction box and multi-classification of objects, which further improves the accuracy of object detection. The experimental result achieves the mAP of 76.2% on the VOC 07 classic dataset, which is 2.7% higher than the Faster R-CNN. On the VOC 12 test, the mAP of 75.6% is improved by 2.5% compared with the Faster R-CNN.