彩色纹理特征抽取和选择算法及其在海洋生物分类中的应用

Texture and color are two important cues in numerous processes of human cognition. It is particularly important to halobios classification. Species recognition and quality classification are mainly relied on both the texture and color characteristics of the creature surface. However most of the current researches on texture classification are based on gray geometric-based features, which makes many objects misclassified because the missing of color contents. This research plans to propose a tensor-based color texture representation method to preserve the spatial structure and color component interaction information of the color image. Based on this model, a multilinear feature extraction algorithm will be proposed for tensor image dimensionality reduction and color texture feature extraction. Two texture feature selection methods will be proposed for local and global texture features respectively, to select the best feature subset for object recognition and image classification. Finally we will explore the application of the proposed methods in halobios classification, including species recognition from halobios image samples acquired by the underwater camera, and scallop classification according to its surface color texture features.

纹理和颜色是人类区分不同事物的重要视觉线索。尤其在海洋生物分类中，生物表面的颜色和纹理特征是分辨其物种类别和品种质量的两个必不可少的因素。但现有的纹理分类研究中大部分是以灰度图像为基础的，这样就导致了颜色这个重要辨别信息的丢失。本课题拟提出图像的彩色纹理张量表示模型，以同时保持彩色图像的本质空间结构信息和色彩分量间的依赖关系；在此基础上提出彩色纹理张量的多重线性特征抽取算法，获得一副彩色图像紧致的纹理特征表示；针对局部纹理特征和全局纹理特征分别提出相应的特征选择算法，用于选择出具有最大相关、最大内部依赖和最小冗余的纹理特征子集用于目标识别和图像分类。最后，我们拟将上述理论模型和算法在海洋生物分类中进行应用探索，主要包括两个方面：一是在养殖海域水下采集的海洋生物图像中对海洋生物的物种进行识别；二是对捕捞的扇贝根据扇贝壳表面彩色纹理特征对其种类和品质进行分类。

水下环境光照变化大，拍摄视角较为单一，成像条件不理想，给图像识别和分类带来了很大难度。针对水下低光照环境导致传统图像特征辨识度不够的问题，本研究提出基于深度学习的水下图像高层特征提取算法和目标识别技术。针对水下样本拍摄视角单一的问题，首先基于仿射变换对样本数据进行增量扩充，模拟不同视角水下目标的形状和表现；在此基础上提出基于深度卷积神经网络的水下图像特征提取算法，通过迁移学习的方式利用有限的水下样本数据训练获得图像的抽象特征表示；最后，我们将上述理论模型和算法在海洋生物分类中进行应用探索，在养殖海域水下采集的海洋生物图像中对海洋生物的物种进行识别。