Temperature Numerical Recognition Based on the Temperature Scale Bar in Infrared Images

This paper proposes a method for small-target temperature numeral recognition in infrared images under complex backgrounds. The approach employs dynamic grayscale thresholding to suppress high-contrast backgrounds, utilizes geometric segmentation with fixed-pixel coordinate localization to precisely identify temperature numeral positions, and implements template matching recognition through a multi-scale Euclidean distance algorithm with a standardized template library. Addressing critical challenges including low image resolution, insufficient contrast, and small target size, the method has been validated on a dataset of 20,000 images, achieving a $98.2 \%$ recognition rate and $41.52 \%$ higher precision than traditional OCR methods. Featuring high accuracy and ultra-low resource requirements, this solution enables real-time deployment on embedded devices for power equipment inspection-providing reliable overheating alerts without dependence on deep learning or large datasets.