Deep Learning-Enhanced UV Fluorescence for Automated Detection of Foreign Bodies in Tilapia Fillets.

Tilapia fillets are widely popular worldwide, but endogenous foreign matter (such as scales and bones) remaining during processing poses potential risks to quality control and food safety. Furthermore, these endogenous foreign objects are difficult to detect through manual or traditional visual inspection methods. This study developed a non-destructive rapid detection method for endogenous foreign bodies in tilapia fillets. After acquiring high-quality images of foreign bodies using a UV fluorescence imaging system (360-370 nm), a U-Net deep learning model was first employed to accurately segment the foreign body regions. Subsequently, color features were extracted from various color models (RGB, HSV, L*a*b*, and YCbCr), and texture features were extracted from images enhanced by principal component analysis (PCA). A support vector machine (SVM) classifier optimized using a genetic algorithm was then constructed. Among these, the model integrating color and local binary pattern (LBP) texture features (Color-LBP-GASVM) performed well, achieving an average accuracy of 95.9% and an overall average F1 score of 96.15% on the test set. The results confirm that combining UV-induced fluorescence imaging with an integrated deep learning and machine learning framework holds great potential for the automatic and reliable detection of endogenous foreign bodies in tilapia fillets.