Image Saliency Detection Based On Weighted K-Nearest Neighbor And Deep Guidance

With the advancement of science and technology,the image information generated by people is increasing rapidly.How to quickly mine the information used by people from the mass of resource data has become a problem that researchers pay close attention to.The continuous exploration of this problem has promoted the rapid development of computer vision.Image saliency object detection(SOD)is a sub-field of computer vision research.As a pre-processing process,it aims to automatically detect and segment the regions of human interest for the image.To get around the problem of improving the performance of image saliency object detection,the dissertation put forward algorithms based on weighted k-nearest neighbor and deep guidance respectively.The specific work is as follows:K-nearest neighbor optimization for SOD algorithm.In order to improve the performance and efficiency of traditional saliency detection algorithm,the dissertation proposes a weighted k-nearest-neighbor linear blending algorithm.First of all,existing saliency detection methods are employed to generate weak saliency maps and obtain training samples.Then,Weighted K-Nearest Neighbor(WKNN)is introduced to learning salient score of samples.WKNN model need no pre-training process,only need selecting k value and computing the distance between the k-nearest neighbors training samples and the testing sample.In order to reduce the influence of selecting k value,linear blending of multi-WKNNs is applied to generate strong saliency maps.Last,multi-scale saliency maps of weak and strong are integrated together to further improve the detection performance.Deep layer guided network for SOD algorithm.In order to solve this problem that how to fuse the features of each layer of the convolutional neural network,a deep layer guided network in which global information from deep layers is progressively transmitted to shallow layers is proposed.Three effective modules are embedded in the network,which are hybrid feature enhancement block,discriminant feature block and salient inference block.Hybrid feature enhancement block receives the feature maps of adjacent layers,and outputs enhanced feature maps which reduce the loss of spatial details and the impact of varying in shape,scale and position of object.Discriminative feature block highlights the consistency in differentlayers from channel and spatial dimensions.The discriminative features of this layer are fused with the features of the previous layer to get the optimized level features in th saliency inference block.The function of these three modules in turn are to enhance the original features of each layer extracted by the feature extraction network,to enhance the discrimination between features and to infer the final saliency map through robust features.The two saliency detection algorithms proposed in this dissertation are evaluated on public saliency datasets.The experimental results show that the weighted k-nearest neighbor linear hybrid algorithm can effectively improve the performance of traditional saliency detection algorithms;deep layer guided network for SOD algorithm can better fuse the features of each layer of the convolutional neural network.The algorithm effectively detect and segment the most attractive regions in the image and its performance is even better than other classic algorithms e under evaluation metrics.