Saliency-based data compression for image sensors

As saliency models have revealed ability to predict where observers fixate during scene exploration, it is of great interest to embed a saliency model into an image sensor to apply a fixation-driven compression. Such perception-driven image sensor can allocate bit-rate budget according to the saliency level of a region. In this paper we present an original implementation of a saliency-based data compression framework to be integrated in image sensors. First, a video-rate compliant, compact saliency model is used to predict salient regions: only compact operators and little memory are required. Second, Haar wavelet-based compression is applied according to the block saliency value. Both saliency computation and data compression are carried out on-the-fly. Our experiments showed that no significant differences between original and compressed videos can visually be perceived.