Research On Visual Saliency Detection And Salient Object Segmentation

With the rapid development of mobile storage technology and Internet technology, the data of global image and video increases rapidly. The images and video information has gone beyond the capacity of computer’s calculating ability, puts forward a new challenge for visual large data processing technology. Humans effortlessly focus attention on important parts of scene from a lot of visual input, and then analyses and processes the data. It is hoped that the intelligent information processing system has the human ability to see from coarse to fine. Visual saliency region detection is to minic the human’s ability for locating import part of scene and salient object segmentation is to minic subsequent process which lay foundation for scene understanding. On the basis of existing research, visual saliency detection and salient object segmentation are deeply studied in this paper.First of all, the local saliency cue of single scale can’t describe the whole salient object in the natural image with complex background. Aiming at this problem, an image pyramid fusion based salient region detection approach is proposed. The image is divided into non-overlapped region which is as basis for saliency analysis, and the feature of segmented regions can be represented by average color in Lab color space. Two saliency cues, including global color contrast and color distribution, are computed in the same scale, and the two cues are fused using nonlinear strategy. Different layers of saliency map are required to adjust to the same scale and the weight of different saliency maps are set to the same value.Secondly, to solve the problem of incomplete detection of low-level visual features, this paper proposes a feature fusion based saliency detection method which combines low-level image feature and high-level prior. A region merging based hierarchical image abstraction is build. On the basis of this, the high-level visual features, including image boundary and image ceneter, and the low-level visual features, including color contrast and FCM-based region color distribution, are computed. And then the features are need to be fused, the progressive and heuristic strategy is used in the same scale while the information entropy based fusion strategy is exploited under different scales. The experimental results show that our method can restrain background effectively.Afterwards, to fully consider the interaction between the different visual cues, this paper proposes a feature combination and learning framework for salient region detection. Four atomic features, including objectness feature, GMM-based color distribution feature, the sparse feature of region which is based image boundary and image center, the global color contrast feature, are extracted from the segmented region. A novel feature combination strategy is used to map a four dimentional vector to a fifteen dimentional vector and a logistic classifier is trained to separate the background and salient region. In the stage of detection, the saliency map is enhanced by multi-scale integration. The experiments result on the public datasets show that the proposed algorithm can generate high quality saliency map.Finally, this paper proposes an automatic object segmentation approach which is based on saliency seed and random work algorithm. During the first stage, use the saliency detection result to generate the inial seed and train a SVM classifier to reassign these seeds labels; During the second stage, combine the saliency seed and random to realize the automatic segmentation of the object; During the third stage, use the mathematical morphology operation to process the initial segmentation result which can remove the isolate point and inner cavity. The qualitative and quantitative analysis is performed on two public data sets, the experiment results show that the proposed method can obtain better segmentation result.