A compact saliency model for video-rate implementation

Salient regions are useful for priority processing of fixated areas in case of limited computational or bandwidth resources. This paper presents an efficient and compact saliency model, compatible with image sensors implementation. In contrast with Itti´s reference model, our model relies on a low-complexity processing consisting of inter-frame difference, spatial contrast, and the central fixation bias. Simple weighted combinations of these factors show better results than Itti´s model. Moreover, a combination using time-varying weights further improves the performance; and results show that spatial saliency should be integrated after motion saliency. Besides, a new criterion compatible with image sensors is used to evaluate the models.