Detection Method from 4K Images Using SSD300 without Retraining
Kei Irie, Kiyoshi Nishikawa · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
In this paper, we propose a method to improve the accuracy of object detection from 4K images using SSD (single shot multibox detector) without retraining. When SSD is used for object detection, the input image will be resized to the size of images used for training the network. Because of this resizing of image, sizes of objects in the image are also affected so that the detection accuracy may possibly decrease. In the proposed method, the image is resized to one of the predefined sizes which is larger than usually used in the applications of SSD. This modification enables the improvement of the detection accuracy when the high-resolution images are inputted. The advantage of the proposed method is that it does not require retraining of the network parameters. The effectiveness of the proposed method is demonstrated by computer simulations.