Automatic Image Segmentation With Superpixels and Image-Level Labels

Abstract

Automatically and ideally segmenting the semantic region of each object in an image will greatly improve the precision and efficiency of subsequent image processing. We propose an automatic image segmentation algorithm based on superpixels and image-level labels. The proposed algorithm consists of three stages. At the stage of superpixel segmentation, we adaptively generate the initial number of superpixels using the minimum spatial distance and the total number of pixels in the image. At the stage of superpixel merging, we define small superpixels and directly merge the most similar superpixel pairs without considering the adjacency, until the number of superpixels equals the number of groupings contained in image-level labels. Furthermore, we add a stage of reclassification of disconnected regions after superpixel merging to enhance the connectivity of segmented regions. On the widely used Microsoft Research Cambridge data set and Berkeley segmentation data set, we demonstrate that our algorithm can produce high-precision image segmentation results compared with the state-of-the-art algorithms.

Weakly Supervised Conditional Random Fields Model for Semantic Segmentation with Image Patches

“…To further improve the results in semantic image segmentation, we will focus on three topics in the future. The first is fusion of color, edge and other information for refined segmentation [47,48], and the second is saliency based extraction of objects from images [49,50]. The third direction is deep learning based image segmentation and object detection, where convolutional neural networks and other models will be explored [51,52], even in combination with the first two topics such as multiscale segmentation and extreme learning machines [53,54].…”

UMMS: Efficient Superpixel Segmentation Driven by a Mixture of Spatially Constrained Uniform Distribution

“…Therefore, superpixels should be more consistent with human visual cognition. It not only extracts visual features conveniently, but also greatly reduces computational complexity of the algorithm [1], [2]. Generally, superpixel segmentation algorithm should meet boundary adherence, pixel similarity, and superpixel regularity, etc.…”

Entropy Rate Superpixel Classification for Automatic Red Lesion Detection in Fundus Images

“…In our previous work [ 28 ], the SLIC algorithm was also applied to extract the superpixels from an image. The SLIC method has been widely used in image processing but has important limitations in terms of accuracy and boundary adherence [ 29 ]. Frequently, it produces some under-segmented superpixels [ 29 ].…”

“…The SLIC method has been widely used in image processing but has important limitations in terms of accuracy and boundary adherence [ 29 ]. Frequently, it produces some under-segmented superpixels [ 29 ]. We propose a novel RL detection method based on superpixel classification which avoids the need for independent blood vessel removal, unlike most RL detection methods.…”

“…ERS is an efficient greedy algorithm which follows a clustering formulation based on an entropy measure [ 27 ]. Under condition of the same initial number of superpixels, the performance comparisons show that the ERS method overcomes the SLIC method in the three standard superpixel segmentation metrics [ 29 ]: Boundary Recall [ 30 ], Under-segmentation Error [ 26 ] and Achievable Segmentation Accuracy [ 30 ]. Boundary Recall measures the level of coincidence between the superpixel segmentation boundary and the ground truth boundary.…”