Coffee flavor is one of the main factors consumers consider when tasting and choosing coffee. Traditional and common methods of assessing coffee aroma are based on expensive equipment or artificial senses. These methods are time-consuming, costly, and require trained evaluators. Therefore, we propose a low-cost and portable coffee recognition method that combines an improved convolutional neural network (CNN) to recognize and classify five different varieties of coffee. The coffee smell is collected through the electronic nose, and an improved CNN classification model with three convolution layers was designed using Leaky-ReLU and Dropout model optimization techniques, which has improved performance compared to commonly used CNN models. Compared with other models used for machine olfaction, the experimental results show that the accuracy of improved CNN model for coffee five classification can reach 84.80%, and the model can effectively classify coffee.