Analysis of existing techniques for image-based recognition of agricultural crops diseases

An analysis of existing methods for processing and identifying the disease of agricultural crops was carried out. Other methods for identifying informative regions, including Fourier transformation, k-means clustering, Histogram Equalization, Scale-Invariant Feature Transform (SIFT), Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG) algorithms, as well as their combinations were reviewed. The studied approaches demonstrated high accuracy in the identification and classification of various plant diseases. The study examined hybrid models, such as, for example, logistic regression with decision trees and Extreme Learning Machines (ELM). The accuracy of the Support Vector Machine (SVM), ELM, and Decision Trees algorithms were compared, convinced, that the importance of choosing the right parameters and fine-tuning to improve accuracy is important. The methods of deep learning such as Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) and their usage in the scope of the image recognition, which are also used for disease recognition, were examined. The recognition accuracy of several CNN models was compared: DenseNet121, MobileNetV2, NASNetMobile and EfficientNetB0, and the last demonstrated the best results. The modification of ready-made architectures, including the architecture of the EfficientNetB0 neural network, has been analyzed as a way of adapting existing models to specific recognition requirements.