结合领域知识的复杂场景城市图像特征提取与识别算法研究

At present, the urban management field has a strong demand for image recognition methods, and has achieved certain success in theory and application. However, there are still some urgent problems such as complex scenes, insufficient sample size, incomplete feature description, weak classification synergy and low computational efficiency. Such issues have constrained further development in this area, and this project is produced for the current needs and shortcomings of urban image management automatic recognition technology. The theory part of the project mainly consists of three part: 1)cascade regression feature point positioning theory based on super-resolution SIFT and PCA shape classification; 2)LBP local feature co-expression theory based on fuzzy analytic dictionary learning, 3)and depth network model based on fast PCA and high-recognition feature optimization. The theories are studied separately, and the local depth identification method and application system for small sample urban management image are developed. The results of this project will make up for the lack of sample capacity and ergodicity of traditional methods combining feature extraction and machine learning, further enrich the theory of existing target recognition, and have important academic value. It can also realize the automatic discovery of anomalies in urban management in complex environments, which is an important way to realize intelligent urban management and has a great application prospect.

目前城市管理领域对图像识别方法需求强烈，已经在理论和应用领域取得了一定的成功，但是仍然存在场景复杂、样本数量不足、特征描述不全、分类协同性弱以及计算效率低下等突出问题，从而制约了该领域的进一步发展。本项目针对当前城市管理图像自动识别技术所存在的需求和不足，从超分辨SIFT和PCA形状分类的级联回归特征点定位理论、基于模糊解析字典学习的LBP局部特征协同表达理论，以及基于快速PCA核的深度网络模型及其高辨识度特征优化理论分别展开研究，研发面向小样本的城市管理图像局部深度辨识方法并应用。本项目研究的成果将弥补特征提取和机器学习相结合的传统方法对样本容量和遍历性的要求的不足，进一步丰富现有的目标识别理论，具有重要的学术价值；还可实现复杂环境下城市管理异常问题的自动发现，是实现智慧化城市管理的重要途径，具有较大的应用前景。

现代社会中，随着科技水平的快速发展，人们在认知周围环境的过程中有很多获取信息的途径，而图像和视频往往会成为最直观、最简单的方式。人类对图像内容的识别主要是通过对图像的像素分布、颜色及纹理等特征的统计，将图像内容所属类别进行正确的分类，如何从图像中获取这些重要的信息成为我们分析和理解周围信息的关键。图像特征提取方法有效解决了这一关键问题。不同类型的图像具有不同的目标特征，因此研究对不同图像类型相适应的特征提取算法具有非常重要的意义。本项目的重点是结合领域知识对图像特征提取与识别算法的研究及应用，针对城市中复杂环境下待识别与检测特征不同，围绕基于深度学习的特征提取算法。项目的主要工作及创新性表现包括以下几个方面：基于深度感知模块的局部特征提取算法以及在特殊场景下的图像识别分类算法研究、基于深度卷积网络的图像增强模型、基于轻量级生成对抗网络框架的图像识别算法以及迁移学习策略的泛化模型、基于注意力感知的长短期记忆网络、基于动态自适应残差网络的分割方法以及它们在复杂场景城市图像特征提取与识别应用中展开研究。