An Algorithm for Selecting the Head and Tail of an Intact Fish in the Overlapping Multi-fish Image for Freshness Detection

The fish freshness detection application assists the public in determining the freshness of fish purchased at the market. The application operates two principal tasks: detecting body parts’ regions of interest (ROI) and classifying freshness. For ROI detection, the You Only Look Once (Yolo) method detects intact fish and their parts, such as heads and tails. Then, a Convolutional Neural Network classifies them for freshness. However, the input image for Yolo may contain fish with arbitrary placement resulting in overlapped and redundant detected parts. Hence, an algorithm to select the appropriate head and tail of an intact fish from the detected parts is required to correctly aggregate the freshness classes of all fish in the image. This study proposes a head and tail selection algorithm using two principal components: the head-tail distance and the intersection over the fish part. The experimental results on 20 overlapping fish images show that the algorithm selects heads and tails with an accuracy of 84.21%. The best weights for both components are 0.6–0.4 to 0.8–0.2.