Improving Efficiency of Automatic Labeling by Image Transformations on CPU and GPU

Abstract

Automatic image labeling is a process which detects and creates labels on various objects of an image. This process is widely used for autonomous vehicles, when original images collected by cameras placed in the car are sent to the artificial neural network (ANN). Improving the efficiency of labeling is vital for further development of such vehicles. In this study, three main aspects of image labeling are analyzed and tested, namely, (i) time required for labeling, (ii) accuracy, and (iii) power consumption.

In this paper, an approach is proposed to improve the efficiency of the serial and parallel implementation of labeling, performed respectively on CPU and GPU. One of transformations used in this approach is converting an original image to its monochrome equivalent. Other transformations, such as sharpening and change of colors intensity, are based on using the image histogram. The testing and validation of image transformations are performed on a test dataset containing frames of proprietary videos collected by an onboard camera. For both serial and parallel labeling, the same ANN is used. Preliminary results of these tests show promising improvements when considering the above-mentioned aspects of automatic labeling.