PERFORMANCE COMPARISON OF DIFFERENT SIZED REGIONS OF INTEREST ON FISH CLASSIFICATION

Öz

In this study, different sized regions of interest were obtained from fish images and these were used for fish species classification. A previously proposed region of interest obtaining method was upgraded in order to acquire wider regions of interest. Depending on general accuracies of classification performances, comparison between these regions of interest was made. According to comparison results the effects of the different sized regions of interest were discussed for classification purposes of fish species. This study was performed by using a database which consists of 1321 fish images. These fish images include fish samples from 16 fish families and 35 fish species. All images were colored in RGB color space. But two different feature sets were extracted for fishes by examining images both in RGB and HSV color spaces. Feature extraction was performed by using a color based method. For each color space, seven statistical features were extracted from each component of the color space. Two feature sets were acquired for each fish sample by combining the extracted statistical features according to color spaces. The obtained feature sets from RGB and HSV color spaces were used separately for classification purposes. Classification was performed according to families and species by using Nearest Neighbor algorithm as classifier. According to classification results, the best performances on general accuracies were achieved as 93.5% and 91% for fish families and species classification respectively.