In the past decade, deep learning made a revolution in many areas of our life bringing the opportunity to build a complex system based directly on data without any handcrafted conditions. Nevertheless, there are a lot of problems which arise when prepare and deploy trained model in production as to collect data, provide labeled training sample, accuracy/performance tradeoff, size of the model, etc. In this report, we propose our solution of the problem of plant diseased recognition based on a deep convolutional siamese network with the triplet loss function. We show how to effectively train a deep neural network on an extremely small dataset and then how to optimize it to be executed on mobile devices with real-time image processing.