A METHODOLOGY OF IMAGES CLASSIFICATION USING PRINCIPAL COMPONENT ANALYSIS AND HIERARCHICAL CLUSTERING

The human eyes detect subtle differences in images. Nowadays technology allows quantifying color, luminosity, format, etc. On the other hand, there are plenty of researches who deal with images in almost different areas including foods, sensory analysis, materials, extrusion process, among many others. Nevertheless, it is not known any statistical method available that can identify similarities among images. Thus, the aim of this work was to develop a methodology to identify similarities within images by applying principal component analyses (PCA) and hierarchical clustering from principal components (HCPC), and to optimize the computational requirements. As example, of the use of this approach, sixteen expanded extrudate images were used (2126 x 2126 pixels/image) from mixtures of sorghum (two genotypes) and coffee powder (0%, 10%, 15%, and 20%) processed by thermoplastic extrusion at two moisture levels (16% and 20%). The images were read and processed using R free software. Each image consisted in three matrixes (Red, Green, and Blue). The data was treated with two computers, a supercomputer with 256 Gb of RAM memory and a regular laptop with 8 Gb of RAM. Prior to apply PCA and HCPC, the images were converted in vectors, resulting in a table with 16 rows x 13559628 columns. Then, the size of the images was gradually reduced to optimize the computational requirements. The results of original and reduced images were similar in PCA and HCPC e.g. more expanded extrudates formed a group, whereas low expanded formed another group. The groups showed same spatial distribution in PCA for all image sizes which corroborated by HCPC. The results allowed the use of a new methodology to classify imagens by applying statistical techniques that could be extremely useful for the scientific community also reducing the computational requirements by using only one regular computer.