Research On Saliency And Co-Saliency Region Detection Algorithm Of RGB-D Image

The image is indispensable in contemporary human life,in order to extract more valuable regions from a large amount of image data,many saliency detection models have been proposed.Many saliency detection models in the past used low-level features such as color and texture to determine the saliency region,this kind of method has certain limitations when the saliency region is similar to the background.However,the models using deep neural networks in the past have complex structures,and the improvement effect is not obvious enough.In view of this problem,this thesis studies the simplicity and effectiveness of the network structure,and proposes to use the simple and effective network model to determine the saliency region.With the progress of the times,the co-saliency detection model between multiple images has been proposed,the previous co-saliency detection model use low-level features to calculate in the whole image,which will reduce the accuracy of the results when the background is complex.In response to this problem,this thesis has studied how to narrow the saliency detection range and better characterize the feature of superpixels,and proposes to use high-level and low-level features to calculate the co-saliency value of superpixels in the object box.The specific work implementation is shown below.This thesis proposes a RGB-D saliency detection model based on multi-scale feature fusion,this model achieves the good experimental result on the basis of simple structure,it consists of three parts,which are multi-scale convolution neural networks,multi-scale deconvolution network,and the optimization.The output image of the last stage of the multi-scale convolution neural network is used as the input of the multi-scale deconvolution network,the deconvolution network has symmetrical structure with convolution network,gradually recover the image size can better preserve the image information and reduce the loss of image information.finally,the side output of the multi-scale convolution network and the output of the deconvolution network are fused using the three-stage fusion method to obtain the final saliency map,this fusion method can amplify the advantages of the upper stage output and the lower stage output to obtain good RGB-D saliency map.This thesis proposes a RGB-D co-saliency detection model based on object detection,which is the first time to propose the use of object detection to locate saliency object in the field of RGB-D co-saliency detection.The model cuts superpixels for each image as the basic unit of processing,merges the optimized initial saliency map,RGB image,and depth image into the model,generates object box through object detection,the superpixels in the box will be the candidate choosing as co-saliency seed superpixels,the color,depth,as well as VGG-16 network second feature are used to calculate the initial co-saliency map,the use of low-level and high-level features together increases the robustness of the model,finally,the optimization method is used to optimize initial co-saliency map,and the final co-saliency map is obtained.The model of RGB-D saliency detection based on multi-scale feature fusion network proposed in this thesis is compared and evaluated with four classical algorithms on three public RGB-D datasets,the experimental results show the performance of this method is better than the classical algorithm.The RGB-D co-saliency detection model proposed in this thesis is tested on standard co-saliency detection dataset RGBD Cosal150,and compared the experimental results with other four mainstream methods,the effect of the model in this thesis is better than other models,and it can be said that it has achieved relatively great results.