The article shows the features of the deep convolutional neural network architecture development for the plant diseases recognition based on their photographic images analysis, laying in the optical wavelength range. An analysis of the structural regularization and complex convolutional filters impact on the recognition quality and training stability is presented. The experiments are carried out with the original lightweight neural network architecture (app. 9.7K trained parameters). Real- life field data is used for training and testing, with photographs taken in adverse conditions: variation in hardware quality, angles, lighting conditions, and complex disorienting backgrounds. The resulting solution achieves the quality of 97.3% (weighted 𝐹1-score, one input image, test dataset).