Abstract: To address the current issues where the adjustment during the drying stage of cigar tobacco leaves relies on manual experience, resulting in inaccurate division and untimely regulation, which leads to poor quality of tobacco leaves after drying. A kind based on the Multi-region Adaptive Feature Fusion architecture and combined with the Convolutional Neural Network and the Long short-term memory neural network was constructed The MAFF-CNN-LSTM cigar tobacco leaf drying stage identification model of Short-Term Memory. Images of cigar tobacco leaves during the drying process were collected using an image acquisition system including an industrial camera and a circular light source. The image data were divided into five drying stages according to the state changes of the tobacco leaves. The sample diversity was expanded through ordinary data enhancement and Mixup mixed enhancement. A comparative experiment was designed to compare the established model with the CNN-LSTM, MAFF-CNN, CNN and ResNet50 models, verifying the validity of the constructed model. The results show that: ① The recognition accuracy rates of the MAFF-CNN-LSTM model for the training set and validation set in the five drying stages are 98.72% and 97.45% respectively; The recognition accuracy rate and average F1 score of the test set reached 98.29% and 98.22% respectively. Compared with the test sets of the CNN-LSTM and MAFF-CNN models, the accuracy rates were increased by 4.40 and 1.53 percentage points respectively. ② The samples with incorrect judgments of the MAFF-CNN-LSTM model are much fewer than those of the other four models, and most of them are in adjacent drying stages, which will not cause a large temperature and humidity difference. This method can provide support for the rapid and accurate identification of the cigar tobacco leaf drying stage and assist in the digital management of the cigar tobacco leaf drying process.