Online monitoring and prediction of moisture content and quality of blueberries during hot air drying based on computer vision and deep learning

• A smart dryer with load cell and CV system was developed. • CV system non-destructively and efficiently tracked the changes of materials. • Res-UNet model accurately performed online image segmentation. • 1D-CNN model precisely predicted the changes of moisture content and shrinkage. In current work, a convective dryer embedded with load cell and computer vision system was applied for online monitoring of blueberries’ physical properties during drying. New image processing method and time-dependent prediction models based on deep learning were developed to extract product features and control the drying process. Experimental results demonstrated that the hot air drying of blueberries follows a deceleration drying pattern. The effective moisture diffusivity ( D eff ) calculated by Fick’s second law ranged from 6.73×10 −11 to 3.58×10 −10 m 2 /s. The color parameters ( L*, a*, b* ), total color difference (Δ E ), and shrinkage rate of blueberries exhibited significant dependence on both drying temperature and duration. Notably, the lowest Δ E value was occurred at 70 °C, while the maximum shrinkage rates were observed at 60 °C. The ResNet-18 model pretrained in ImageNet could precisely perform semantic segmentation on blueberry images. The convolutional neural network (CNN) model was adopted to effectively predict the changes of moisture content and shrinkage rate of blueberries at different drying temperatures with acceptable R 2 and RMSE, i.e. 0.984, 3.66e −2 , and 0.997, 9.72e −3 , respectively. This study tested an intelligent-enabled prototype dryer with the potential for adaptive regulation integrated quality strategies and showed superior performance in blueberries dying.