Objective: The performance of current machine vision is far from that of human vision. Simulating human visual mechanism is an effective way to improve the existed algorithm. Human visual system can detect objects with high acuity and focus attention on region relevant to current visual task. These advantages all owe to visual attention mechanism. Human accept attention by making a serious of eye movements. There are two forms of eye movement: saccades and microsaccades. 1) In saccades stage, human eyes aim to find candidate object so it makes sharply shifts in the whole field of view. 2)While candidates are identified as target, eyes will make a series of dense tiny movements that is called microsaccades around the target for the purpose of intensify objects and inhibit noises. Continuous microsaccades will lead to visual fading and the eye movement will switch to the stage of saccades to find new objects. The integration of saccades and microsaccades contribute to the quickly and efficiently performance of human vision system. Motivate by above facts, this paper presents a novel saliency detection framework by simulating microsaccades and visual fading. A positive feedback loop is constructed that focus on fixation area and intensify objects to make saturation of visual perception that leads to visual fading. In which, multiple random sampling of the gaze area is used to simulate the behavior of microsaccades, and RVFL (Random Vector Functional Link Networks) is utilized to simulate the human neural system to produce binary visual stimulus. The proposed framework is data-driven totally, need not any prior knowledge and labeled samples.
Method: Firstly, the conventional saliency detection methods could be used to produce a variety of saliency map. We group these saliency maps to an integrated saliency map to simulate multi-channel visual perception. The integrated saliency map can be thresholded further to form an initial fixation area. Followed multiple random sampling could be executed from the pixels in the fixation and non-fixation area. Then ensemble of RVFL (Random Vector Functional Link Networks) is trained on-line by those samples of pixel. And then the model of RVFL could be used to classify image pixels to obtain a new fixation area (binary area). For the new fixation area and non-fixation area, iterations of " sampling – learning (modeling) - pixel classification" could be performed on-line. If the fixation area were unchanged in the iteration, it indicates that the perception is saturated and the iteration should be terminated. If taking binary result of pixel classification as a kind of visual stimulation, the output of multiple visual stimuli could be accumulated to generate new image saliency map. And the last binary result of pixel classification in positive feedback loop could be regard as foreground of segmentation.
Result: Three popular image databases SED2, MSRA10K and ECSSD were chosen to evaluate the performance of our algorithm. They total contain 11100 nature images with different salient objects and scenes. Every image in the dataset was finely labeled manually for the purpose of saliency detection and image segmentation. Five other models were compared, include the state-of-the-art or closely related to our approach: BL,RBD,SF,GS and MR. P-R curve, F-measure and MAE was used to illustrate the performance of the algorithm in six algorithms on three databases. Experimental results show that our method has the best performance in SED2 (two objects) and MSRA10K(single object). Our method is inferior to BL and very close to RBD in the ECSSD (complex scene and multi-object) database, while better than the rest compared algorithms. It is also shown that the performance of BL,RBD,SF,GS and MR. can be improved effectively by adding learning-based positive feedback in SED2 database.
Experimental images illustrate that the new method is more consistent with the visual saliency map of human perception by positive feedback and accumulating visual stimulation. From the view of qualitative evaluation, it is clearly that the binary result detected by our method is closer to the Ground truth than others. The positive feedback iteration could be saturated quickly, and the running time of algorithm is not significantly increased. It can be treated as an effective post-processing modular, which could improve the performance of the conventional saliency detection algorithm.
Conclusion: This paper proposes a novel saliency region detection method base on machine learning and positive feedback of perception. Motivated by human visual system, we construct a framework using RVFL to process visual information from coarse to fine, to form saliency map and extract salient objects. Our algorithm is data-driven totally and need not any prior knowledge compared with the existed algorithms. Experiments on several standard image databases show that our method not only improves the performance of the conventional saliency detection algorithms, but also segments object successfully in different scenes.