基于高光谱图像的巴沙鱼煎制状态识别方法研究

Research on a Method for Identifying the Frying State of Basa Fish Based on Hyperspectral Imaging

摘要: 为实现巴沙鱼不同煎制状态的快速、客观与无损判别，本文提出一种利用高光谱成像技术构建巴沙鱼煎制状态智能识别方法。搭建起巴沙鱼煎制状态分类依据，将巴沙鱼煎制过程划分为生、半熟、熟和过熟四种状态。通过高光谱相机采集不同煎制阶段巴沙鱼样本的空间–光谱三维数据，并对原始数据进行黑白校正与固定空间裁剪，构建标准化高光谱数据集。在特征建模方面，设计了一种CsTsrpModel分类模型，引入相对位置编码增强模型对空间–光谱位置信息的感知能力以及稀疏注意力机制优化注意力计算过程，提高模型对重要特征的聚焦能力。结果表明，CsTsrpModel模型在巴沙鱼煎制状态感知中取得了最佳性能，其准确率、精确率、召回率和F1分数分别达到94.83%、94.94%、94.83%和94.79%。研究结果验证了基于高光谱成像与深度学习方法在巴沙鱼煎制状态识别中的可行性与有效性。

Abstract: To achieve rapid, objective, and nondestructive discrimination of different frying states of Basa fish (Pangasius bocourti), this study proposes an intelligent identification method based on hyperspectral imaging technology. A classification framework for frying states was established, in which the frying process was divided into four categories: raw, half-cooked, fully cooked, and overcooked. Spatial-spectral three-dimensional data of samples at different frying stages were acquired using a hyperspectral camera. The raw data were subjected to black-white calibration and fixed spatial cropping to construct a standardized hyperspectral dataset. In terms of feature modeling, a classification model named CsTsrpModel was developed. Relative positional encoding was introduced to enhance the model’s perception of spatial-spectral positional information, and a sparse attention mechanism was incorporated to optimize the attention computation process, thereby improving the model’s ability to focus on critical features. The results demonstrate that the CsTsrpModel achieved the best performance in frying state recognition, with accuracy, precision, recall, and F1-score reaching 94.83%, 94.94%, 94.83%, and 94.79%, respectively. These findings validate the feasibility and effectiveness of integrating hyperspectral imaging and deep learning for frying state identification of basa fish.