深度学习进展及其在图像处理领域的应用

[摘要] 深度学习一般通过3种方式进行：有监督学习、无监督学习和混合深度学习。以“无监督或生成式特征学习”以及“有监督特征学习和分类”为例，讨论了深度学习及其在图像处理等领域的进展及未来可能的研究方向。认为深度学习打破了传统机器学习和信号处理技术普遍基于浅层结构的局限。得益于相关非凸优化等问题的逐步解决，深度学习已经在图像处理等领域取得了一些突破性的进展。

[关键词] 深度学习；图像处理；分层结构

[Abstract] Deep learning methods are usually divided into three different categories, including supervised learning, unsupervised learning and mixed-structured learning. In this paper, "unsupervised or generative feature learning" and "supervised feature learning and classification" are taken for examples to illustrate the research advances in deep learning, and several promising research lines in image processing. Deep learning technique has overcome some limitations of shallow structures commonly used in traditional machine learning and signal processing. Due to the progress in some aspects, such as relevant nonconvex optimization, deep learning has achieved remarkable development.

[Keywords] deep learning; image processing; hierarchical structure