CFD-augmented deep learning for accurate non-destructive prediction of squid drying behavior using RGB images

초록

This study presents a multimodal deep learning framework for non-destructive prediction of squid drying characteristics under varying drying strategies and temperatures. Drying behavior was characterized by analyzing moisture ratio (MR) and drying rate curves, which revealed faster moisture removal under intermittent drying (ID) compared to continuous drying (CD), particularly at 50 degrees C. Computational fluid dynamics (CFD) simulations accurately captured internal moisture and temperature distributions (RMSE <0.061 for MR; <2.12 degrees C for temperature), providing critical image data that compensated for the limitations of real image datasets and surface-based measurements. Convolutional neural network (CNN) models were individually trained using RGB images, infrared (IR) thermal images, and CFD-generated contours. The CFD-trained model achieved the highest classification accuracy (95.7 %), demonstrating the advantage of internal dynamic representation. However, the RGB-only model performed less reliably (87.8 %), particularly during later drying stages when surface changes diminished. To overcome this limitation, we trained a multimodal CNN using RGB, IR, and CFD data, and deployed it using only RGB images. The resulting model achieved >99 % accuracy in classifying drying stages across all conditions, including challenging ID scenarios. This improvement is attributed to crossmodal feature abstraction and knowledge transfer enabled by CFD-augmented training, allowing internal drying dynamics to be inferred from surface-level RGB inputs. This approach effectively bridges the gap between limited experimental imaging and rich, physically-informed simulation data. Compared to traditional sensor-based or unimodal image-based methods, this framework offers a more accurate and scalable solution for visual monitoring of drying processes, particularly in scenarios where internal measurements are inaccessible or costprohibitive.

키워드