Saliency Detection Algorithm Based On Multi-scale Image

Salient object detection is one of the core contents in the field of computer vision,with broad research prospects and application field.In recent decades,the developing era of salient object detection brought substantial and great algorithms.The data-driven(bottom-up)salient detection algorithms mostly use the contrast ratio between pixels as the saliency of the pixel according to the characteristics of the image in the data set.Such algorithms do not require pre-training,and the algorithms are faster in general with a high accuracy when dealing with regular image data sets.By contrast,task-driven(top-down)saliency detection algorithms require a large amount of data sets and time for pre-training,but their ability of dealing with various data sets is excellent.Based on the fact that the images of low-resolution contain little information,prominent features of salient area,While the information of high-resolution images is abundant and redundant.The following improved methods are proposed.Since the saliency areas of natural image locate at image center in general.The salient region can be obtained by calculating the distance between each pixel and the boundary.Based on the above theory,many algorithms with higher accuracy or fast velocity are proposed.However,for improving the efficiency and performance of the algorithm,further researches are still needed to do to obtain faster and more accurate saliency detection algorithm.To handle the above problems,a saliency detection algorithm based on multi-scale minimum barrier and gradient fusion is proposed.First,the multi-scale image is formed by multiple sampling,and the saliency detection algorithm based on the improved minimum barrier is introduced to calculate the saliency of the processed image.Second,during the process of the algorithm,gradient analysis is performed on the image and background image.The both images are fused to eliminate the problem of fuzzy region.The saliency detection algorithm with prior knowledge is not universal,which is result in weakly managing plentiful and complex images.Therefore,this paper introduces the convolutional neural network to deal with the problem of salient detection.A model is proposed for extracting multi-view image features and self-learning strategies of multi-scale feature fusion.In this model,in order to achieve multi-view feature extraction,we adopt the dilated convolution which is outstanding in tasks such as image segmentation.The dilated convolution is placed in the same convolution layer to capture the context of multiple receptive fields in multi-level convolution.The model will learn how to fuse the feature images which character with multi-scale and multi-view.After that,the fused image is mapped to a saliency predicted image.In addition,we use the image of top layer to guide the prediction of the underlying image to clearly segment the edges.Experimental results demonstrate that the proposed model reaches the level of the state-of-the-art algorithms.