基于可学习Nestrov稀疏编码器的深层模型研究

The key to the study of Deep Hierarchical Models (DHMs) with data sparsity prior lies in designing efficient and learnable Sparse Encoders (SEs). However, the existing SEs are unable to balance well between efficiency and adaptability, which has seriously restricted the pratical use of related DHMs. To solve these issues, we study the following DHMs in this project. Firstly, a learnable deep hierarchical SE with low complexity is developed from a first order proximal algorithm for sparse optimization, which is based on the Nesterov’s acclerated gradient decsent method, thus being named as the Learnable Nesterov’s SE (LNSE). Secondly, to address the high complexity issue of end-to-end DHMs with existing learnable deep hierarchical SEs, LNSE based end-to-end DHMs are proposed, especially for image super-resolution, video foreground/background segmentation and image clustering tasks. Thirdly, in order to conquer the high complexity and weak adaptability problem of the Convolutional Sparse Coding (CSC) model built with existing non-learnable SEs, the LNSE based CSC model is presented. We further investigate the related unsupervised image feature learning network. This project will establish a new way for designing DHMs with data sparsity prior, and the proposed algorithms can promote the development of intelligent information technologies for marine and shipping, which indicates the significance of this project in terms of both theory and practice.

基于数据稀疏性的深层模型研究的关键在于设计快速且可学习的稀疏编码器，而现有的稀疏编码器无法兼顾效率和适应性，这严重影响了相关深层模型的实用性能。针对该问题，本项目的研究内容包括：（1）从基于Nesterov快速梯度下降法的一阶近似稀疏优化算法出发，给出一种低复杂度、可学习的深层稀疏编码器——可学习的Nesterov稀疏编码器；（2）针对基于现有可学习稀疏编码器的端到端深层模型计算复杂度高的问题，面向图像超分辨率、视频前背景分割和图像聚类，研究基于可学习Nesterov稀疏编码器的端到端深层模型；（3）针对基于现有不可学习稀疏编码器的卷积稀疏编码模型计算复杂度高且适应性弱的问题，构建基于可学习Nesterov稀疏编码器的卷积稀疏编码模型，研究基于该模型的无监督图像特征学习网络。本项目将为基于数据稀疏性的深层模型设计提供一种新思路，相关技术可服务于海洋和航运智能信息化，具有重要的理论意义和应用价值。

基于数据稀疏性的深层模型研究的关键在于设计快速且可学习的稀疏编码器，而现有的稀疏编码器无法兼顾效率和适应性，这严重影响了相关深层模型的实用性能。针对该问题，本项目的研究内容包括：（1）从基于Nesterov快速梯度下降法的一阶近似稀疏优化算法出发，给出一种低复杂度、可学习的深层稀疏编码器——可学习的Nesterov稀疏编码器；（2）针对现有端到端深层模型计算复杂度高的问题，面向图像处理具体问题，研究高效的端到端深层模型；（3）研究弱监督或无监督图像特征学习网络。本项目将为基于数据稀疏性的深层模型设计提供一种新思路，相关技术可促进深度学习的工业应用，具有重要的理论意义和应用价值。