Handheld UV-C Fluorescence Imaging with Cascaded Denoising and Deep Learning for Escherichia coli Concentration Classification on Diverse Food-Contact Surfaces

Rapid, non-destructive assessment of Escherichia coli contamination levels on food-contact surfaces is important for food safety monitoring and sanitation verification. This study developed a deep-learning framework integrating UV-C fluorescence imaging and cascaded denoising to classify predefined E. coli concentration levels on four food-contact surfaces. A SafetySpect CSI-D+ handheld system acquired fluorescence images at four concentration levels: Control (0), 106, 107, and 108 CFU/mL, on stainless steel, white rubber, white plastic, and black plastic. Five image-denoising operators (wavelet thresholding, NLM, unsharp masking, Wiener filtering, and RL-TV) were evaluated in sequential and parallel configurations. Eight deep-learning classifiers, including five YOLO11-cls variants, EfficientNetB7, ConvNeXtBase, and ViT, were benchmarked using consistent evaluation protocols. Classification performance varied with surface type and denoising strategy. The sequential Wavelet → Wiener → RL-TV cascade achieved the highest observed mean test accuracy (94.18%), with surface-specific accuracies of 97.57%, 97.57%, 94.44%, and 87.15% for stainless steel, white rubber, black plastic, and white plastic, respectively. On the most challenging surface (white plastic), the cascade improved accuracy by 11.46 percentage points over the best single operator. YOLO11-based models generally outperformed ViT, although the highest-performing architecture varied by surface. Most misclassifications occurred primarily between 106 and 107 CFU/mL, while the 108 CFU/mL class was consistently classified accurately. Overall, the ordered cascaded denoising can improve robust E. coli concentration classification across diverse food-contact surfaces, providing a promising foundation for rapid, non-destructive food safety inspection. The results also highlight the importance of surface-specific preprocessing and deep learning architectures for robust performance across diverse materials.