Indonesian local coffee, widely known in the global market, is vulnerable to being counterfeited with other cheaper coffee products. Therefore, we need technology to identify the types of local Indonesian coffee. One of the non-destructive methods for identifying coffee products is computer vision. This study aimed to develop a computer vision method to classify three types of local Indonesian Arabica coffee beans. Those are Gayo Aceh, Kintamani Bali, and Toraja Tongkonan using three types of pre- trained convolutional neural network (CNN), namely GoogLeNet, SqueezeNet, and Inception-v3. Sensitivity analysis was carried out by varying the optimizer, i.e. SGDm, Adam, and RMSProp, and varying the learning rate, which included 0.00005 and 0.0001. Each type of coffee bean used 500 image data for training and validation with a ratio of 70% and 30% and 100 image data for testing. The results show that GoogLeNet, SqueezeNet, and Inception-v3 can achieve up to 100% in validation accuracy value and up to 99.67% in testing accuracy. From the results of this study, the pre-trained CNN model, SqueezeNet with an SGDm optimizer and a learning rate of 0.00005, is highly recommended to classify local Indonesian coffee beans. This is because it has the highest validation accuracy, the highest testing accuracy, the simplest CNN structure, the fastest training time, and the most stable training and validation charts.