The computational analysis of wheat images to identify wheat varieties and quality has wide applications in agriculture and production. This paper presents an approach to the analysis and classification of images of wheat samples obtained by the method of crystallization with additives. In tests 3 concentration and 4 times for each concentration were used, such that each type of wheat was characterized by 12 images. We used the images obtained for 5 classes. All the images have similar visual characteristics, that makes it difficult to use statistical methods of analysis. The multifractal spectrum obtained by calculating the local density function was used as a classifying feature. The classification was performed on a set of 60 wheat images corresponding to 5 different samples (classes) by various machine learning methods such as linear regression, naive Bayesian classifier, support vector machine, and random forest. In some cases, to reduce the dimension of the feature space the method of principal components was applied. To identify the relationships between wheat samples obtained at different concentrations, 3 different clustering methods were used. The classification results showed that the multifractal spectrum as classifying sign and using the random forest method in combination with the principal component analysis allow identifying wheat samples obtained by crystallization with additives, being the highest average classification accuracy is 74%.