A Teacher-Student Model for Unsupervised Foreground Segmentation

• Main Authors: Hung-LunTai, 戴宏倫
• Other Authors: Min-Chun Hu
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/23w874

碩士 === 國立成功大學 === 資訊工程學系 === 107 === In recent years, deep learning has been popular in both researches and applications. There are pixelwise classification with deep learning research, for example, semantic segmentation[2][3][4] and instance segmentation[5][6][7]. These methods feature the capability of precisely recognizing every pixel and classifying objects in a given image. In the other words, it can distinguish the contour of object in an image. In past, before deep learning training, two crucial works have to be conducted, data preprocessing and data annotation. Data annotation costs much especially in the part of semantic segmentation which is labeling the contour of objects. Aimed to solve the high cost from handcraft data labeling, we present an end-to-end teacher-student model of unsupervised foreground segmentation in this study. In addition, we apply a Dual Encoder Attention Network (DEANet) as teacher network in order to explore the similarity structure between object features from encoder. DEANet network reconstructs the foreground and background of image respectively, and generates the mask of object without data labeling. Next, we apply the Hierarchical Network as student network to learn how teacher network works to generate masks. We find that student network in this study gets more accuracy details on object contour. After experiments, we get results with 0.83 and 0.94 of Jaccard Similarity on University of Oxford Flower Dataset and Stanford Cars Dataset, which are pretty good results. The result is shown in Figure 2 as below.