High-throughput and objective fillet color profiling and defect assessment are crucial for quality control and pricing benchmarks in the fish processing industry. Our research group has developed and tested a hand-held, smart device (FilletCam) that employs computer vision and artificial intelligence to enable real-time fillet color scoring. The initial research prototype demonstrated satisfactory performance with consistent accuracy and repeatability in color measurements; however, its functionality was limited to color-related quality assessments. Furthermore, the convolutional neural network (CNN) model was trained on a relatively smaller dataset collected under controlled lighting, which may affect its generalizability and adaptability to environments with variable or non-uniform lighting, such as fish processing plants and retail stores. This study aims to enhance the existing prototype by expanding its capabilities beyond color scoring to include fillet defect detection and color uniformity assessment. To improve model robustness, we are expanding and diversifying the training dataset with images captured under different lighting conditions and annotated for common fillet defects, including gaping, melanosis, blood spots, and scale residue. The dataset will be split into training (70%), validation (20%), and test (10%) subsets, and a custom model for defect detection and color profiling will be trained using Roboflow (Roboflow, Inc., Des Moines, Iowa, USA). The refined CNN model will be deployed on the upgraded FilletCam 2.0 hardware, and its performance will be evaluated in terms of mean average precision and F1 score. Predicted results will be compared against expert-annotated ground truth data, and findings will be presented.
Keywords: Quality control; Fillet defect; Fish processing; Precision aquaculture; Artificial Intelligence; Computer vision