Huang J.-YDing J.-JHuang P.-C.JIAN-JIUN DING2022-04-252022-04-25202122195491https://www.scopus.com/inward/record.uri?eid=2-s2.0-85123199311&doi=10.23919%2fEUSIPCO54536.2021.9616319&partnerID=40&md5=c484f2fa3503aafed66c5b7cef4d2390https://scholars.lib.ntu.edu.tw/handle/123456789/607205Most conventional segmentation methods are superpixel-based. Recently, the convolutional network (CNN) has been adopted in image segmentation. However, most existing CNN-based segmentation algorithms are pixel-wise. Due to the irregular shape and the non-fixed size of superpixels, it is hard to apply superpixels into the CNN architecture directly. In this work, several ideas are proposed to solve this problem. Instead of applying the whole image as the input directly, we apply a square patch that contains only two superpixels as the input of the CNN. Also, instead of generating the segmentation result directly, the output of the CNN is whether the two superpixels should be merged. The proposed algorithm integrates the merits of conventional superpixel-based methods, feature-based methods, and CNN-based methods. Simulations show that the proposed algorithm can achieve very high accurate segmentation results and outperform state-of-the-art methods in all metrics. ? 2021 European Signal Processing Conference. All rights reserved.Computer visionDeep learningImage segmentationSuperpixel mergingMergingSuperpixelsConvolutional networksImages segmentationsIrregular shapeNetwork-basedSegmentation algorithmsSegmentation methodsSegmentation resultsSuper pixels[SDGs]SDG3conference paper10.23919/EUSIPCO54536.2021.96163192-s2.0-85123199311