Research On Image Analysis And Attribution Algorithm Based On Super-pixel

As an important information carrier,image has the characteristics of high dimensionality and large amount of information.With the development of computer vision algorithms,the ability of machines to recognize and understand images continues to improve.However,in some professional fields such as medicine and biology,human experts need machines to provide the attribution analysis of the local information of the image while providing the prediction results.Therefore,the problem of image analysis and attribution has attracted more and more attention.Researchers propose a series of superpixel segmentation algorithms to parse the image into local conceptual blocks,and obtain the importance of different analysis units by backpropagating the gradient method.However,these methods have problems such as single conceptual block shape,limitation of attribution objects,rule constraints and relaxation.In view of the shortcomings of previous image analysis and attribution algorithms,the main work carried out in this dissertation is as follows:On the one hand,for the problem of irregular concept blocks that are difficult to solve in image analysis tasks,we proposes a graph theory-based superpixel segmentation algorithm called Extract and Merge(EAM).Through the proposed power window structure,feature extraction is performed on the input image,and the concept of regional attributes is defined.The image features are divided into central region features and peripheral region features.Through modeling of the extracted power windows,an attractive competition mechanism is constructed.The power windows are merged from the central region to the peripheral rigion to form super-pixel blocks.On the other hand,in view of the limitations in the previous attribution algorithms,we propose a black-box attribution algorithms based on random sampling.The algorithm performs random walk sampling by moving the position of the proposed dynamic circles.According to the change of the model’s prediction results,we shrink the area of the dynamic circles and obtain the image analysis block with super-pixels.Through the proposed attribution algorithm,the dissertation draws the conclusion that "more complex networks need less information to make a specific prediction".We evaluates the proposed method on multiple public benchmark data sets.Experiments show that the proposed method can solve the above problems well and show excellent performance in the quantitative results.