Classification of Tuna Meat Grade Quality Based on Color Space Using Wavelet and k-Nearest Neighbor Algorithm

• Main Authors: Putra, I Gede Sujana Eka, Darma Putra, I Ketut Gede, Sudarma, Made, Kompiang Oka Sudana, Anak Agung
• Format: Conference Proceeding
• Language: English
• Subjects: Classification algorithms; Feature extraction; Fish; grade; Image color analysis; Image preprocessing; k-Nearest Neighbors; Shape; Training; wavelet
• ISSN: 2831-400X
• DOI: 10.1109/ICSGTEIS60500.2023.10424189

Tuna products are one of Indonesia's leading export commodity products. Accuracy in determining the quality grade of tuna is necessary to ensure food safety and product quality. Several cases of rejection of Indonesian fishery products by the United States and cases of food poisoning show a lack of food safety management in Indonesia. Determination of fish quality grade was done manually by identifying the eyes condition, gills, and meat color, taking a sample of tuna meat using a spike, which is the manual grading process causing human error. Previous studies had been carried out to identify the freshness and quality of fish using odor profiles, and color profiles from color sensors, measuring fish freshness and quality based on fish eye images. According to Robert DiGregorio, five parameters determine the grade quality of tuna, namely freshness, fish size and shape, meat color, texture, and fat content. Fish grades are grouped into grade 1, grade 2+, grade 2, and grade 3. This study aims to determine the grade quality of tuna meat based on color space through image preprocessing, dataset training, and classification. Image preprocessing consists of image cropping, converting images from RGB to HSV, and feature extraction using a wavelet. The training phase uses the k-NN algorithm using k=4 based on the number of classes. The result shows the correlation coefficient between grades of feature extraction using wavelet Symlet better than Haar. Classification of 65 images test dataset using Symlet wavelet and k-NN has a better accuracy of 81.8% compared to the Haar wavelet k-NN with an accuracy of 80.3%.