Research On Structure-Sensitive Superpixel Image Segmentation Method

Superpixel image segmentation plays a very important role in image processing.Superpixels can express local region information well and maintain the boundary,which is a feature that is not possessed by a single pixel in traditional image segmentation.Image preprocessing for superpixels is more conducive to image processing and understanding.With the emphasis on the superpixel segmentation method,more and more superpixel segmentation methods have emerged.Among them,in view of the inconsistency of the internal information of superpixels generated by over-segmentation,the researchers proposed a structure-sensitive superpixel segmentation method.The structure-sensitive superpixel segmentation method generates small superpixels in the structure-sensitive region to ensure the superpixel internal information is consistent,and generates large superpixels in the sparsely structured region to ensure the operation speed.We added a boundary term based on the original MSLIC to ensure the boundary fit of the structure-sensitive superpixel.Local area pixel distribution constraints are incorporated into the original SEEDS method to optimize the structure-sensitive superpixels that generate shape rules and boundary fits.Mainly done the following two aspects:(1)A structure-sensitive superpixel image segmentation method with boundary terms is proposed for the part where the structure-sensitive superpixel does not fit well.This method considers the structural sensitivity to segment the region,introduces the boundary feature,calculates the possibility of the pixel on the real image boundary through the boundary term,and then redistributes and optimizes the superpixel boundary,so that the generated super pixel boundary is as much as possible.Fits real image boundaries while ensuring small superpixels generated in structure-sensitive areas and large superpixels generated in sparse areas.The Berkeley dataset was used to analyze and compare the segmentation images through eight evaluation indicators.The proposed method has good performance in performance such as under-segmentation error rate,recall rateand accuracy.(2)Aiming at the problem that the SEEDS algorithm has fast segmentation speed but the segmentation of superpixel shape is irregular,a superpixel segmentation method based on local region pixel distribution is proposed.The method is based on the energy function of color term and boundary term in the original SEEDS algorithm.The local area pixel distribution metric is added,and the boundary image is generated by performing boundary detection on the image.The geodesic distance is used to further constrain the superpixels that generate the shape rule according to the boundary image.At the same time,the distance measurement of the geodesic can ensure the structural sensitivity of the image.Generate small superpixels in structure-sensitive regions,generate large superpixels in sparse regions,and ensure the content integrity of the generated superpixels in local regions.The experiment uses the images in the Berkeley dataset to evaluate and analyze the segmented images through eight evaluation indicators.It proves that the proposed method has good performance in achieving segmentation accuracy,compactness and under-segmentation error rate.