Research On Visual Salient Object Detection In Complex Scene

From the 80 s of last century,scientists have come to explore the posibility of computer simulating human eye vision mechanism to recent years,visual salient object detection models have applied in the mainstream business product,visual saliency has long been touchstone to validate artificial intelligence theory.Saliency object detection is to detect and split the distinctive object that draw human attention.Salient object detecction methods can be classified into two categories.The superpixel pre-segmentation based region-level saliency detection method and the FCN based pixel-level saliency detection method.The region-level saliency detection method can efficiently keep the object contour,while performs poor in complex scene.The pixel-level saliency detection method can detect the salient object in most case.While it is limited by its network struture,which leads to an inaccuracy segmentation result.In the light of above problems,the following algorithms will be discussed.Aiming at the problem that pixel-level salient object detection algorithms cannot segment object accurately,a modified FCN based network structure for salient object detection is proposed.Except for the dilated convolutions,we introduce short connections into this structure to optimize the network structure.A joint loss function is defined to guide the network training process.A visualization contrastive analysis shows our method can effectively improve the detection effect.Aiming at the problem that regoin-level salient object detection algorithms preform unsatisfying in complex scene,a background refinement and saliency diffusion based method is proposed.Objectness feature is introuduced to segement the initial background area and improve background prior based methods performance.Moreover,a statistics on feature distribution is utilized for weight contrasting to refine the saliency ranking process.To further improve the salient object detection accuracy,a multi-model saliency fusion and refinement method is proposed.This method combines the advantage of pixellevel and region-level saliency detection algorithms to optimize the saliency map.Experiments on several well-designed benchmark datasets for complex scene demonstrate that our method performs well under complex scene.