Image Clipping Strategy of Object Detection for Super Resolution Image in Low Resource Environment

Abstract

Nowadays, deep learning based object detection algorithms have been widely discussed in various application fields. Some super resolution images are not fully supported by the deep learning network structure in low resource hardware devices. An example is the embedded hardware installed on satellite with the super resolution image as input. Traditional methods usually clip the large scale image into small multiple block images, and merge the detect results from clipped images. However, there is still a lack of corresponding strategy to analyze the relationship between detection efficient, including cost time, detection accuracies and the size of block images. To this end, this paper proposes a strategy of image clipping for super resolution images on low resource hardware environment of object detection algorithm. In methodology, we consider the relationship among the number of chip cores, the number of image targets, image input resolution, image block numbers and propose an evaluation method for the optimal efficiency of object detection on low resource hardware. In experiments, the classic Haar detection algorithm runs on TI6678 chips with 8 CPU cores shared with 2.0 G RAM. It shows that the utilization efficiency of object detection algorithm in chip is related to the ratio of segmentation, the number of partitions and the number of CPU cores used. The exploration in this work is also helpful for researchers and developers optimize other algorithms, such as Fast R-CNN, YOLO, on other similar low resource environment in the future.

L. Huang and X. Zhang—The authors contribute equally to this paper.