Frost detection and thickness estimation using a transformer-based semantic segmentation model

Abstract

Frost accumulation on evaporator surfaces of the refrigeration systems reduces heat transfer efficiency, increases power consumption, and disrupts cooling uniformity. Efficient defrosting based on frost detection and quantification is critical for maintaining the performance and energy efficiency of refrigeration systems. Although several methods with different strategies have been proposed in the literature, accurate frost detection and reliable thickness estimation remain critical challenges. This study presents a novel frost detection and quantification pipeline based on a Transformer-based segmentation model and a data-driven pixel-based regression method. High-resolution RGB images of heat exchangers were captured under controlled refrigeration conditions and preprocessed to create a labeled dataset for model training and validation. The proposed method employs the SegFormer architecture for semantic segmentation, generating precise binary masks that effectively distinguish frost-covered regions from the background. The segmentation achieved a mean accuracy of 93% and a mean Intersection-over-Union (IoU) of 0.875, demonstrating strong generalization across varying frost densities and illumination conditions. For frost thickness estimation, using binary masks from SegFormer, three datadriven formulations were analyzed for quantitative mapping between the normalized frost pixel values (dimensionless) and the experimentally measured frost thickness (mm). The best performance was obtained with the quadratic formulation, achieving an R2 of 0.930 and an error margin of 11.98%. Additionally, comparative evaluation with previous image-processing methods using the same dataset revealed substantial improvement in results.