Image Saliency Detection Via A Local And Global Method Based On Deep Neural Network

People usually give priority attention to the most important and conspicuous object regions when they face the scene within the scope of complex visual area, and then they can catch up the most important and conspicuous object regions in the short time by using their own visual information processing system from complex scenarios. Saliency detection serving as a preprocessing step, can efficiently focus on the interesting image regions related to the current task and broadly facilitates computer vision applications such as segmentation, image classification, Video tracking, and compression, to name a few. It determine the probability of each pixel in the image belongs to the interested area through a certain algorithm or model simulating the human visual system, according to different criteria for evaluation each pixel can be assigned to the specified category, the final output is the scene saliency map.Deep neural networks have achieved state-of-the-art results in image classification, object detection and scene parsing. The success stems from the express abilities and capacity of deep architectures that facilitates learning complex features and models to account for interacted relationships directly from training examples. It broke the traditional characteristics of artificial selection in the past, greatly promote the development of a significant target detection, Since DNNs mainly take image patches as inputs, they tend to fail in capturing long range label dependencies for scene parsing as well as saliency detection. The network becomes more mature as the image contexts is more impeccable, we usually use a recurrent conventional neural network to consider large contexts.In this paper we summarize the predecessors of saliency detection scientific researches and combine with the state-of-the-art deep artificial neural researches and technologies, we put forward a method of saliency detection via a local and global method based on Deep Neural Network. In this paper, the proposed model is divided into three stages:Local stage, Global phase, local and global convergence phase. In the local stage, we use a local estimation network to study the characteristics of the local image block used to determine the probability of the local each pixel in the image block belonging to saliency regions, thus getting the local saliency map. And then we use a refinement method of the geodesic object proposal to extract a set of object segments, thus incorporating object level concepts into local estimation to enhance the spatial consistency of local saliency maps. In the global phase, we use global search network to obtain contrast and geometry information of the entire image. The contrast and geometry information can use as the global features of the image to determine the probability of each pixel in the image belongs to the interested area. To make up the loss foreground information, we propose the propagation optimization algorithms based on the color-distance information, then the refinement saliency map is obtained. In the local and global convergence phase, we take advantages of the least square method to study the optimal weights, thus making a linear combination of the local stage and global stage, finally we can gain the final saliency map.The proposed model is evaluated and compared with ten state-of-the-art methods on five public standard salient object detection databases. Experimental results show that the proposed method comfortably outperforms other state-of-the-art methods in terms of PR-curve, F-measure and visual quality for the local network, global network and the overall network. The method we proposed can detect the saliency object accurately and effectively.