Abstract: In high-voltage disconnector state recognition, a single image modality generally provides limited information, resulting in low recognition accuracy. To address this problem, a gradient-based image fusion method was proposed to fuse the infrared and visible images of high-voltage disconnectors, followed by state recognition using a lightweight MobileNetV2 network. First, the infrared and visible images were decomposed into high- and low-frequency subband images using the non-subsampled shearlet transform (NSST). The high-frequency subband images were then fused using the multi-scale morphological gradient (MSMG) and weighted local energy (WLE), whereas the low-frequency subband images were fused using the guided filter (GF) and sum of modified Laplacian (SML), thereby achieving local fusion. Subsequently, the fused high- and low-frequency subband images were reconstructed using inverse NSST to obtain the final fused image and achieve global fusion. Finally, a lightweight MobileNetV2 network was employed to recognize the states of high-voltage disconnectors. Experimental results showed that the proposed method achieved an average recognition accuracy of 93.9%, significantly outperforming several commonly used methods.