基于深度学习的高光谱图像红肉品质检测理论与技术

Red meat quality detecting technology based on hyperspectra has seen some achievements in recent years. Compared with traditional artificial sensory evaluation methods and standard physic-chemical methods, it is more objective, accurate, and real-time. In this project, the research on using depth-learning technology to detect red meat quality is conducted. Through hyperspectral image technology, a large number of image information reflecting external sensory quality (color, shape, texture, etc.)and spectral information reflecting internal physic-chemical quality ( pH value, TVB-N, drip loss, fat content, etc. ) of red meat are collected to establish a hyperspectral database of red meat quality. On the basis of a large number of hyperspectral image data, a stacked automatic encoder algorithm can be used to reduce the dimensionality of the hyperspectral data, and the convolutional neural network as a classifier, the hyperspectral data cube in the rectangle centered on the pixel to be classified is used as the input of the convolutional neural network,perform aerial spectrum joint classification to obtain the final classification result. The classification is designed to reveal the relationship between all qualities and corresponding physic-chemical indexes of red meat and its hyperspectral images, a quantitative detection model and a quality comprehensive evaluation model for physic-chemical indexes of red meat quality are established. The project explores ways of rapid non-destructive testing of the red meat quality, and the research results are conducive to improve the quality of our meat products.

近些年基于高光谱的红肉品质的检测技术取得了一定的成果,比起传统的人工感官评定法和标准的理化方法，它具备更客观、精准和实时等优点。本课题开展利用深度学习技术来检测红肉类品质的研究，通过高光谱图像技术大量采集反应红肉外部感官品质(颜色、形状、纹理等)的图像信息和反应内部理化品质(PH值、TVB-N、滴水损失、脂肪含量等)的光谱信息，构建红肉品质高光谱数据库。在大量高光谱图像数据的基础上，利用堆叠自动编码机对高光谱数据进行光谱维上的降维，然后将卷积神经网络作为分类器，将以待分类像元为中心的矩形内的高光谱数据立方体作为卷积神经网络的输入，进行空谱联合分类，获得最终的分类结果，建立红肉品质理化指标的定量检测模型和品质综合评判模型。本课题为红肉类品质快速无损检测和分级提供了思路，具有重要意义。

本课题采用基于深度学习的人工智能技术来检测鉴定红肉类品质。通过采集大量红肉的高光谱图像样本以及样本对应的物理内部理化(PH值、TVB-N、滴水损失、脂肪含量等）指标来构建红肉品质高光谱数据库。项目提出了若干的深度学习模型来解决高光谱数值预测品质的问题，其中，针对高光谱图像和可见光图像信息融合的问题，提出了基于注意力机制和对抗生成网络的高光谱图像融合模型；针对图像分辨率不高问题，提出了高光谱图像的超分辨率模型；针对高光谱图像光谱带之间相似度和维度较高，使用原始数据进行分类会造成光谱信息难以挖掘甚至维数灾难的问题，提出基于空谱融合与协同表示的高光谱图像分类模型。本项目成果将形成一套准确有效的高精度红肉品质数据分析方法，促进光谱分析技术在食品检测领域的进一步发展。