Multi-Scale Deep Subspace Clustering With Discriminative Learning

Deep subspace clustering methods have achieved impressive performance compared with other algorithms. However, most existing suffer from the following problems: 1) they only consider global features but neglect local in self-expressiveness learning; 2) discriminative information of each coefficient matrix; 3) ignore useful long-range dependencies and positional feature representation learning. To solve these problems, this paper, we propose a novel multi-scale deep learning (MDSCDL) to obtain high-quality matrix. Specifically, MDSCDL bridges multiple fully-connection layers between encoder decoder learn matrices features, representing more comprehensive relationship among data. By modeling interdependencies matrices, adaptively assigns weights for matrix fuses them convolution operation. Moreover, increase power, introduces coordinate attention mechanism extract Extensive experiments on face object datasets shown superiority several state-of-the-art methods.