Infrared imaging detection of industrial ethylene leakage faces the problems of insufficient edge feature representation, shallow local-global information interaction, and poor adaptability of feature weights, resulting in limited detection accuracy for low-concentration leaks and limited robustness in complex backgrounds. This paper proposes a new method for infrared ethylene leakage detection based on the adaptive multiscale transform domain, integrating the Edge Attention Module (EAM), Cross-Attention Module (CAM), and feature fusion gating: the EAM strengthens the edge features of leak contours via the Sobel gradient operator, the CAM establishes an information interaction mechanism of “global localization-local focusing”, and the feature fusion gating dynamically balances the contributions of the two types of features. Experiments on the EL infrared ethylene leakage dataset show that the accuracy of the proposed method is significantly better than that of traditional detection methods; it also maintains the advantage of lightweight, meeting the real-time monitoring needs of industrial edge devices.