Compact Saliency Model and Architectures for Image Sensors

Abstract

In this paper we present an original implementation of a compact saliency model for image sensors. The saliency model combines two features: motion and the central fixation bias. Its implementation was designed for low complexity: it relies on compact operators and requires merely about one frame memory. On-the-fly computation allows for low latency processing of "scanline" readout of image sensors. The results show that the proposed model is suitable for video-rate computation and exhibits better performance than the state-of-the-art model in predicting the human fixation. Moreover, a variant of the proposed model further reduce required memory by a factor of 256 while providing results similar to the state-of-the-art algorithm.