A novel object detection algorithm based on enhanced R-FCN and SVM

Abstract

Object detection is an extremely important part of computer vision. However, the object detection result of R-FCN is not good enough in terms of speed and accuracy. In this paper, a novel architecture called Enhanced R-FCN (ER-FCN) is proposed for object detection. Two improvements are presented in ER-FCN. Firstly, novel anchor boxes, 3 scales with box areas of 5122, 2562 and 1282 pixels, and 3 aspect ratios of 0.618:1, 1:1 and 1:0.618, are designed to suit the different scales object detection in RPN. Hence, the performance of object localization and detection speed are increased. Secondly, since the softmax classifier is not optimal to deal with the binary classification problem, a Whale Optimization Algorithm based on support vector machine, termed WOA-SVM, is introduced to improve the accuracy of classification. Extensive experimental results on PASCAL VOC 2007 and PASCAL VOC 2012 datasets show that the mean average precision of ER-FCN is improved by 3.9% compared with that of R-FCN.