Study on texture classification of cantaloupe based on machine vision

Abstract. Texture is one of important appearance qualities of cantaloupes. Grading of appearance quality is mainly processed by a manual operation for a long time, the grading operation is a heavy workload which is laborious and inefficient. An intelligent method regarding texture feature extraction and feature selection was developed to sort different texture type of cantaloupe based on machine vision. A total of 60 cantaloupes of Xizhoumi No.17 were purchased from local market, and then they were manually classified into three categories by an expert. A machine vision system was designed and utilized. Texture features were extracted from the region of interest (ROI) which on the top middle of a cantaloupe (ROI: 300x300 pixels), using different extracting methods, including Gray level co-occurrence matrix, Gray level difference statistics, Gauss Markov Random Field, Differential Box-counting and Uniform Local Binary Pattern. The corresponding number of texture features extracted by these methods were 8, 4, 12, 1 and 59 respectively. Feature selection algorithms of Sequential Forward Search, Genetic Algorithm and Maximum Relevance Minimum Redundancy were applied for feature reduction, the number of selected features were 35, 36 and 17 respectively. The classification accuracy was 96.67%, 95.00% and 92.50% respectively which calculated by the classification model based on support vector machine (SVM) classifier with 4-fold cross validation. Results showed that Sequential Forward Search has the best classification accuracy with minimum time consumption. It can be concluded that multiply texture features combined feature selection can be used in classification of cantaloupes with different exterior quality.