The on-line detection of moisture content in green tea processing is important for stabilizing product quality. To address the shortcomings of traditional manual assessment such as low accuracy and poor consistency, this study proposed the use of on-line NIR spectroscopy to establish a rapid, accurate and quantitative method for the detection of moisture content in green tea processing. 415 samples of green tea leaves from multiple seasons and batches were collected as research materials, and the NIR spectral signals of green tea leaves were collected in real time by an on-line NIR spectrometer. The true moisture content of green tea leaves was subsequently determined by the traditional drying method. Preprocessing methods such as Savitzky-Golay (SG), Standard Normal Variate (SNV), Multiplicative Scatter Correction (MSC) and Detrending were used to eliminate the noise in raw spectra. The genetic algorithms (GA) and Competitive Adaptive Reweighted Sampling (CARS) were compared for the selection of moisture-related characteristic wavelengths. Partial Least Squares regression (PLS) algorithm was then employed to develop PLS, GA-PLS and CARS-PLS quantitative detection models for moisture content. The results showed that the spectral preprocessing could reduce the noise interference in the raw signal, and significantly improve the prediction accuracy of the models, among which SNV achieved the optimal performance. Both GA and CARS could effectively reduce the data dimensionality, among which CARS achieved the optimal prediction performance, the prediction accuracy of the proposed SNV-CARS-PLS model for the samples in prediction set performed with Rp=0.9402 and RMSEP=1.57%. The research established an on-line method for the detection of moisture content of green tea during fixation, which could provide real-time quantitative feedback on the quality condition of the fixation process, and also provide an important basis for the intelligent processing of green tea.

The traditional recognition algorithm is prone to miss detection targets in the complex tea garden environment, and it is difficult to satisfy the requirement for tea bud recognition accuracy and efficiency. In this study, the YOLOv7 model was developed to improve tea bud recognition accuracy for some extreme tea garden scenarios. In the improved model, a lightweight MobileNetV3 network is adopted to replace the original backbone network, which reduces the size of the model and improves detection efficiency. The convolutional block attention module is introduced to enhance the attention to the features of small and occluded tea buds, suppressing the interference of the complex tea garden environment on tea bud recognition and strengthening the feature extraction capability of the recognition model. Moreover, to further improve recognition accuracy for dense and occlusive scenarios, the soft non-maximum suppression strategy is integrated into the recognition model. Experimental results show that the improved YOLOv7 model has the precision, recall, and mean average precision (mAP) values of 88.3%, 87.4%, and 88.5%, respectively. Compared with the Faster R-CNN, SSD, and original YOLOv7 algorithms, the mAP of the improved YOLOv7 model is increased by 7.4, 7.9, and 3.9 percentage points, respectively, and its recognition speed is also promoted by 94.9%, 46.2%, and 16.9%. The proposed model can rapidly and accurately identify the tea buds in multiple complex tea garden scenarios - such as dense distribution, being close to the background color, and mutual occlusion - with high generalization and robustness, which can provide theoretical and technical support for the recognition of tea-picking robots.