A Study On Saliency Detection Based On Deep Convolutional Neural Network And Image Segmentation

This dissertation mainly studies the two fundamental issues of com-puter vision:saliency detection and image segmentation.Faced with large number of images introduced by the exploding technologies,saliency de-tection can be applied to help extract the useful imformation in advance,which plays a large role in making further image analysis and understand-ing more efficient and intelligent.For natural image segmentation,it has been already applied in practice,while in the medical field,because of the complexity of White Blood Cell(WBC),nowadays there has not been an effective and fast method for WBC segmentation.However,WBC analy-sis provides important basis of diagnosis,so realizing intelligent analysis and its automation would be a major breakthrough.Based on above background research,this dissertation focuses on salien-cy detection and WBC segmentation,including saliency detection via con-volutional networks based on feature map selection,WBC location based on task-driven saliency detection,and WBC segmentation based on WBC location.More specifically,1.Considering that for different vision tasks,contribution values of large number of feature maps in convolutional neural networks(CNN)def-initely appear different,this dissertation chooses a pre-trained CNN as an example to take a deep study on the feature maps of each layer.The con-tribution is evaluated by match value defined in this dissertation,then the corresponding feature maps with more contribution would be selected.The dissertation proposes a new saliency detection system,where the selected features are fed into a newly designed CNN to generate the final saliency map.So the proposed method is able to exclude the useless noisy features.In addition,the number of parameters could be reduced to a great extent,and the computing cost can be cut down as well.In this way,the saliency detection can be implemented simply and effectively.2.From the research on the exsisting algorithms,it can be found that the most common cause of inaccurate WBC segmentation is the ill-effect from red blood cell(RBC)in the uneven dying and illumination circum-stances.Therefore,this dissertation considers locating WBC approximate-ly before the segmentation procedure,which can avoid the above proplems,significantly enhancing segmentation accuracy.The dissertation explores WBC location algorithm based on data-driven saliency detection,choos-es the edge density and color contrast cues to measure sliding windows.The idea is motivated by the observation that WBC region always has the deepest color,and its texture is more complex.The final location window is obtained by merging the multi-window with higher scores together.Be-sides,experiments based on Cella/Vision and Jiashan dataset are conducted to verify the validity of the proposed algorithm.3.GrabCut algorithm with human interaction has the higher segmenta-tion accuracy,but limited in practical application,because it need a person to get initialization.In this dissertation,the located sub-image takes the ini-tialization role,therefore,accurate results and automation can be achieved at the same time.Moreover,during the segmentation,the replacement op-eration on nucleus pixels is designed to reduce the contrast between cyto-plasm and nucleus,improving the GrabCut performance further,which all have been confirmed through experiments.