高湿高雾环境下红茶发酵图像实时采集系统研制

为解决高湿高雾的红茶发酵场景下，无法实时采集图像的问题，该研究基于加热除雾防潮装置与工业相机系统，设计了高湿高雾环境下红茶发酵叶图像实时采集系统。系统主要由工业相机、镜头、加热除雾防潮装置、面光源、固定支架、控制软件与电脑组成。利用该系统采集单芽、叶片、茎、一芽二叶的发酵进程图像，通过图像处理技术，可以有效提取发酵叶的颜色特征信息；利用茶叶感官审评、发酵叶色泽与气味变化的方法，将采集的1 564张图像分为发酵轻、偏轻、适度、过度4个带有红茶品质标签的数据集，数据集按照7∶3的比例划分为训练集和测试集，训练集通过MobileNetV3 + Vision Transformer技术搭建深度学习模型，模型预测准确率达到95.68%；开发的可视化红茶发酵程度判别软件，可以实现红茶发酵程度有效预测与结果输出。研究结果表明，设计的图像实时采集系统可为红茶发酵程度智能化判别装备研发提供数据基础和技术支持。

Real-time image acquisition system for black tea fermentation under high humidity conditions

Abstract: Fermentation is one of the key steps for the flavor in black tea processing. However, the traditional evaluation of fermentation grade can often be time-consuming, labor-intensive, and inaccurate at present, depending on the subjective sensory for the color observing and the flavor smelling. Since the change of color dominates the fermentation degree, machine vision can be expected to accurately and timely detect the optimum condition of fermentation in black tea processing. Nevertheless, the real-time image acquisition is still lacking so far, due mainly to the high humidity of processing workshop during tea fermentation. Herein, a real-time image acquisition system was designed for the black tea fermentation under the high humidity conditions. The high-efficient system consisted of an industrial camera, lens, defogging and moisture-proof device, surface light source, fixed bracket, vision software platform, and industrial computer. Specifically, a heating and mist elimination device was used as an upper and lower detachable aluminum alloy shell (length × width × height, 120 × 65 × 68 mm) and two end caps. Among them, three components were soldered, including an opening (50 × 45 mm) in the center of one end cover of the shell, the electrode strip resistance (9-10 Ω), and the wires. The high-temperature resistant insulating transparent glue was used to fix the electrodes on the short side of the ITO coated glass (length × width × height, 55 × 50 ×1.1 mm), where the glass surface resistance with the electrode strip was about 10 Ω. As such, a heating and defogging device was formed to fix the glass on the end cover of the central window. The 12.4 MP color CMOS vision system was equipped with the industrial cameras and the gigabit ethernet interface for 1Gbps bandwidth. The cameras were synchronized with the hardware and software triggers. C mount type was selected to connect the low-distortion 8 mm and 6 MP lens for the machine vision cameras. The specific procedure was as follows. The industrial camera was firstly fixed on one L-shaped bracket, and another was on the sidewall of the aluminum alloy shell by two screws, where the two brackets were interlocked, and the lens was 23 mm away from the ITO coated glass. Three outlet holes with the diameters of 20, 15, and 15 mm were opened on the other end cover, which was used for the outlet of the camera net cable, power cable, and heating power cable, respectively. The outlet and screw fastening positions were then sealed with the high-temperature resistant insulating transparent adhesive. The image acquisition device was fixed for the surface light source on the machine vision bracket. The curtains of the tea fermentation room were drawn to prevent the change of sunlight intensity, then the power of all devices was turned on to adjust the voltage of 6-6.5V in the heating and defogging device. The vision software platform was run for the white balance of the camera once, and then turned off the automatic exposure mode to select the image saving format and continuous trigger for the acquisition mode and the image acquisition interval. The camera was mounted over the surface of the tea, and then manually focused at a distance of about 145 mm. The results demonstrated that the machine vision system was efficient to capture the change of color of tea shoots, leaves, buds, and stem during tea fermentation. 1564 fermentation leaves images were selected to combine with the sensory evaluation and the variations in the color and smell of fermentation leaves. Four datasets were classified with the quality labels: light-, semi-, well-, and over- fermented. The datasets was then divided into the training set and test set, according to the ratio of 7:3. The training set was used to establish the deep learning model by the MobileNetV3 and Vision Transformer. The test set was used to evaluate the classification performance of the model, where the average prediction accuracy reached up to 95.68%. More importantly, a visualization software was also conducted to predict the fermentation grade of black tea using the deep learning. Consequently, the image acquisition system can be expected to acquire the real-time images of fermented leaves under the high humidity and fog environment. The finding can also provide a strong reference for the fermentation evaluation during black tea processing.