Active infrared thermography enables large-area, rapid, and non-contact detection, yet the lateral thermal diffusion-induced edge blurring poses challenges for quantitative defect evaluation. This study investigates quantitative detection in alumina ceramic materials containing flat-bottomed blind-hole defects through long-pulse infrared thermographic nondestructive testing. A novel “triple-height width” temporal feature extraction method was proposed and integrated with a data-driven deep learning model based on Long Short-Term Memory networks. The algorithm was first trained and tested on the simulated dataset, and then its effectiveness was verified through the experimental dataset. Results demonstrate that the proposed method achieves a mean absolute error below 0.05 mm (relative error <2%) for dimeters ranging from 1 to 10 mm in simulated conditions, while maintaining a 5% average relative error in experimental validations. These outcomes highlight the method's high precision and robust generalization capability for defect dimensional quantification, which provides important technical support for the engineering application of infrared thermal imaging technology.