With the advent of resistant varieties and hybrids of tomato, vegetable growers are less likely to encounter diseases on tomatoes. To prevent crop loss, it is important to identify unhealthy tomato leaves and separate them from healthy leaves. Early detection of tomato diseases through deep learning can help decrease the adverse effects of diseases, and also helps surmount the drawbacks of continuous human monitoring. This study examined the performance of modern convolutional neural network classification architectures, such as ResNet18 with a standard algorithm, as well as using optimization parameters and InceptionV3, on 11,000 images of tomato leaves for the classification of tomato diseases. The accuracy of training in Inception V3 was 80.9%, and the accuracy of validation was 71.8%. ResNet architecture training with the momentum parameter demonstrated a high recognition result with 97.7% accuracy. The recognition results were compared using the optimization parameter with values 0.5, 0.7 and 0.9. The influence of the optimization parameter on the quality of training was observed. It can be concluded that using the momentum optimizer with a higher value gives the best results by minimizing fluctuations and increasing accuracy.