Measurement and Analysis of Detecting Fish Freshness Levels Using Deep Learning Method

Abstract

Subjective and objective tests used to determine the fish deterioration process require specialized skills and time, making them inefficient for use by the general public in markets. The quality of fish products in markets is not always guaranteed, so consumers must determine their suitability. Deep learning can be used to analyze images and automatically and accurately detect the freshness of fish. This study aims to evaluate the efficiency of deep learning models in detecting fish freshness and implementing them into an Android application for public use. "Image datasets and pH tests were collected as references for the postmortem phase over a 24-hour period, with hourly checks on three fish species (Rachycentron canadum, Trachinotus blochi, and Lates calcarifer). Data were classified into three classes, pre-rigor/fresh, rigor mortis/semi-fresh, and post-rigor/not fresh. The dataset was divided using the 10-fold cross-validation method and analyzed using YOLOv5 and Faster R-CNN algorithms. The study results showed that YOLOv5 had higher average values for each metric compared to Faster R-CNN. Dataset 8 in YOLOv5 showed precision of 99.4%, recall of 98.1%, f1-score of 98.7%, accuracy of 99.3%, and mAP of 99.3%. The YOLOv5 model for dataset 8 was selected for implementation in the Android application due to its high metric values. This application effectively provides information on fish freshness detection and confidence scores.

Keywords

• fish freshness
• yolov5
• faster-rcnn
• object detection
• android appliaction
• ph