Weakly Supervised Image Semantic Segmentation Based On Superpixels And Neighborhood Rough Set

Image semantic segmentation is a core research issue of pixel granularity in the field of computer vision,which aims at assigning each pixel in an image to one of the pre-defined semantic labels.As a hotspot in academia and industry,image semantic segmentation is not only the key to image understanding,but also provides powerful technical support for practical application scenarios such as autonomous driving,robot perception,video surveillance,augmented reality,m edical diagnosis and so on.At present,the training and learning of image semantic segmentation has the disadvantages of less pixel-level labeling data and high cost of fine-labeled,which restricts the further improvement of precision and the scalability of the model in image semantic segmentation.Therefore,weakly supervised image semantic segmentation based on image-level labels has attracted wide attention from researchers at home and abroad.Compared to pixel-level labels,image-level labels can be quickly and largely obtained from multimedia sharing websites,and can be labeled efficiently and accurately.However,image-level labels contain less guidance information compared to pixel-level labels and other weakly supervised labels.Furthermore,the weakly supervised image semantic segmentation based on image-level labels suffers from the greatest challenges and difficulties.In this paper,we focus on the key and difficult issues in weakly supervised image semantic segmentation based on image-level labels,including the following four research contents:(1)Aiming at the superpixel segmentation can not adaptively generate the initial superpixel number adaptively according to the characteristics of image itself and the quality of superpixel segmentatio n is not high,an adaptive high-precision superpixel segmentation method is proposed.The proposed method mainly focuses on the adaptive generation of initial superpixel number,the detection and re-classification of isolated pixels,and the detection and re-segmentation of under-segmented superpixels.Firstly,according to the minimum peak value of the L color component histogram corresponding to the salient region,an adaptive initial superpixel number generation scheme is proposed,which can solve the de ficiency that the initial superpixel number needs artificial enumeration.Secondly,the large number of isolated point and regions is proved,and a new scheme of isolated pixel re-classification relying only on LAB color space is proposed.Finally,based on the standard deviation and mean of the L color components,an under-segmentation superpixel detection and re-segmentation scheme is proposed.(2)Aiming at the problem of inaccurate edge of SLIC superpixel segmentation and high computational complexity i n superpixel merging,an image segmentation method based on SLIC and neighborhood rough set is proposed.The proposed method mainly focuses on SLIC-based high-precision superpixel segmentation and superpixel merging based on neighborhood rough set.Firstly,the initial superpixel number is generated by indirectly determining the initial grid step size of SLIC superpixel segmentation;the number of isolated pixels is reduced by increasing the compactness factor in SLIC superpixel segmentation to improve the boundary adherence of SLIC superpixel segmentation.In the superpixel merging stage based on neighborhood rough set,information table and neighborhood granule composed of superpixels and their features are constructed.Then,the neighborhood rough set is introduced into the superpixel merging,and the superpixels within the neighborhood granule are merged on the basis of the spatial adjacency.(3)Aiming at the problem of automatic determination of termination conditions in image segmentation,an automatic image segmentation method based on superpixels and image-level labels is proposed.The method mainly includes three key steps: superpixel segmentation based on spatial distance,superpixel merging based on image-level labels,and the reclassification of disconnected regions.Firstly,based on the minimum side length corresponding to the circumscribed rectangle of the salient region,a scheme of initial superpixel number generation based on the minimum spatial distance is proposed.Secondly,a small noise superpixel in an image is defined,and a detection and reprocessing module of small noise superpixel is proposed based on the color and spatial distance between the noise superpixel and the adjacent superpixels;without considering the adjacency between superpixels,a superpixel merging method for automatically determining the termination condition is proposed based on the number of image-level labels.Finally,in order to enhance the connectivity of the segmented candidate regions,an area-based disconnected region detection and reclassification scheme is proposed.(4)Aiming at the problem that the candidate region construction and the inference precision are not high in semantic label inference,a weakly supervised semantic segmentation method based on can didate region and neighborhood classifier is proposed.The method can achieve semantic label inference and prediction at the candidate region level.Firstly,in the training phase,a high-precision candidate region generation scheme is proposed based on the multiple of the label number contained in image-level labels,which can automatically determine the termination conditions;based on the number of labels corresponding to the inferred semantic label,a label inference order determining scheme based on maximum dissimilarity is proposed;based on the multiple of the number of image-level labels and the number of images corresponding to the inferred semantic label,a semantic label inference scheme based on the most similar neighborhood granule is proposed;a decision table composed of inferred candidate regions and their features and semantic labels is constructed;based on the discriminative features of attribute reduction,a weakly supervised semantic segmentation model based on neighborhood classifiers is proposed.Secondly,in the testing phase,the candidate regions obtained by superpixel segmentation and superpixel merging are used as basic processing units to implement label prediction of testing images.