Saliency Object Detection Research Based On Absorbing Markov Chain

A significant goal is that humans use vision to quickly extract the most prominent and important objects from complex life scenes.Compared with human vision,computer visual inspection is still more challenging.The current classical algorithms and deep learning algorithms for computer vision detection still have problems to be solved in the field of image saliency target detection,such as incomplete image detection,darkness,or inconsistency,ambiguity,or target boundary for dark images or foreground and background colors close to the image.Not clear.Aiming at the above problems,a saliency detection method based on Markov absorption chain is proposed.First,stretch the contrast of adjacent pixels,and based on the global neighboring pixel contrast,to develop an optimal function to enhance the image contrast,to expand the difference between the foreground and the background,to highlight the complete saliency goal;secondly,add different in the image The high-frequency component of the degree,using Gaussian difference and corner detection to enhance the image details;then,constructing the series and parallel filters,and performing weight-adaptive multi-structure morphology processing on the image to make the target boundary clear and the curve smooth;finally,A bidirectional Markov absorption chain is constructed on the superpixel image to generate a significant image of foreground a priori and background prior,and the two significant images are merged by the geodesic distance method to generate a final saliency map.Using the MSRA10 K dataset and ECSSD dataset,compared with the current nine popular target detection algorithms,the experimental results show that the final saliency map of this method can highlight the foreground,suppress the cluttered background,and the saliency target is complete and the image The details are clear,the boundaries are smooth and continuous,and both the accuracy and the recall rate are superior to other algorithms.The thesis has 23 pictures,2 tables,and 60 references.