Traditional mechanical fault diagnosis methods for gearboxes are limited by sensor installation constraints and interference from environmental noise, making it difficult to achieve health monitoring of complex or precision mechanical systems. Infrared thermal imaging technology, which uses non-contact detection of surface temperature distribution, offers a novel perspective for fault feature extraction. However, its engineering application is hindered by transmission delays and high computational burdens due to the high-resolution of thermal images (typically 1440 × 1080 pixels). This paper proposes a dual-channel compressed diagnostic framework that integrates wavelet-based global frequency-domain features with region of interest (ROI) local enhancement characteristics. This constructs a lightweight convolutional neural network (CNN) for efficient diagnosis. Experimental results demonstrate that, compared to traditional diagnostic methods, the proposed approach improves processing speed by over 3.2 × with less than 1% accuracy loss. This provides reliable technical support for real-time equipment condition monitoring in fields such as wind power and rail transportation.