On Improving the Accuracy of Object Detection for High Resolution Images Based on SSD

Kei Irie, Yicheng Qiu, Kiyoshi Nishikawa

Abstract: In this paper, we consider improving the accuracy of object detection using SSD (Single Shot multibox Detector). SSD is well known that it can execute the searching of region candidate and classifying the objects in a single process. One of the problems of the SSD is that if the object is not larger than a certain size, the accuracy will decrease. To solve this problem, we propose a method to improve the accuracy of the object detection by extending a conventional method. Although the proposed method applies the SSD after the segmentation of the image as in the conventional one, the difference is that SSD is performed using only partial information from the network. This allows us to detect small objects that could not be detected using the standard SSD. Then, region metrics are corrected using the result of detection without image segmentation and the result after image segmentation. The effectiveness of the proposed method is shown through the computer simulations.