基于高光谱显微成像技术的樱桃番茄损伤检测分类

Damage Detection and Classification in Cherry Tomatoes Based on Hyperspectral Microscopic Imaging

摘要: 水果品质检测是现代农业生产与流通中的关键环节。樱桃番茄作为高营养价值的小型果蔬，在运输过程中极易因振动产生机械损伤。此类隐性损伤不仅影响果实外观品质，还可能加速腐烂变质，进而影响食用品质并危及消费者健康。传统宏观检测方法难以有效识别隐性损伤，也无法进行早期预警。为此，本文提出了一种基于高光谱显微成像系统(HMI)与深度学习相结合的樱桃番茄隐性损伤检测方法。本研究通过自建振动试验台，以模拟实际运输过程中樱桃番茄因振动产生的隐性损伤。随后，利用HMI系统采集正常与损伤樱桃番茄的HMI数据，构建了用于樱桃番茄隐性损伤检测算法训练与评估的数据集。为探究HMI数据多维感知信息在樱桃番茄损伤检测中的可行性，分别针对一维、二维与三维HMI数据构建相应深度学习模型。针对一维光谱数据，基于VGG16网络框架构建1DCNN网络，以充分发掘光谱数据中的深层特征，提升隐性损伤样本的识别准确性；针对二维图像数据，基于ResNet18网络构建樱桃番茄隐性损伤分类模型，通过引入特征金字塔网络(FPN)对基础模型进行改进，以增强对多尺度结构的感知能力；针对HMI三维空间数据，在基础3DCNN框架中引入通道与空间注意力机制，以提高复杂数据的处理能力，避免深层次的关键特征丢失。实验结果表明，改进的3DCNN模型(SC-3DCNN)的准确率为0.98，各类别的精确度超过0.98，召回率超过0.96，F1分数超过0.97，优于其他模型。总之，基于显微高光谱数据的樱桃番茄隐性损伤检测模型能够准确识别运输过程中的隐性损伤，同时验证了显微高光谱三维数据在检测中的优势。

Abstract: Fruit quality inspection is a critical component in modern agricultural production and distribution. Cherry tomatoes, as small fruits with high nutritional value, are highly susceptible to mechanical damage caused by vibration during transportation. Such latent damage not only compromises the fruit’s visual quality but may also accelerate decay and spoilage, thereby affecting edible quality and posing risks to consumer health. Traditional macroscopic inspection methods struggle to effectively identify such latent damage and cannot provide early warnings. Therefore, this paper proposes a latent damage detection method for cherry tomatoes based on the integration of hyperspectral microscopic imaging (HMI) and deep learning. A custom-built vibration test platform was constructed to simulate latent damage caused by vibration during actual transportation. Subsequently, HMI data from both undamaged and damaged cherry tomatoes were collected using the HMI system to construct a dataset for training and evaluating the latent damage detection algorithm. To investigate the feasibility of utilizing multidimensional perceptual information from HMI data for cherry tomato damage detection, corresponding deep learning models were developed for one-dimensional, two-dimensional, and three-dimensional HMI data. For one-dimensional spectral data, a 1DCNN network based on the VGG16 framework was constructed to fully extract deep features from spectral data and enhance the recognition accuracy of latent damage samples. For two-dimensional image data, a cherry tomato latent damage classification model based on the ResNet18 network was developed. The base model was improved by incorporating a Feature Pyramid Network (FPN) to enhance its perception of multi-scale structures. For HMI three-dimensional spatial data, channel and spatial attention mechanisms were introduced into the basic 3DCNN framework to improve complex data processing capabilities and prevent loss of deep-level key features. Experimental results show that the improved 3DCNN model (SC-3DCNN) achieves an accuracy rate of 0.98, with precision exceeding 0.98 for all categories, recall exceeding 0.96, and an F1 score exceeding 0.97, outperforming other models. In summary, the microscopic hyperspectral-based detection model for latent damage in cherry tomatoes accurately identifies hidden injuries during transportation, while validating the advantages of microscopic hyperspectral 3D data in detection applications.