To address the challenges of small targets, missed detections, and low image resolution in transmission line fault detection, a multi-scale feature fusion-based algorithm for transmission line fault detection is proposed. To obtain input features at different scales, images are first fed into a pre-trained shared residual network. Then, a feature fusion attention mechanism is used to learn salient features at different scales, integrating detailed information from large-scale feature maps and contextual information from small-scale feature maps. Additionally, inspired by dense spatial pyramid pooling in semantic segmentation, a multi-scale feature fusion dense pyramid is constructed to further enhance feature extraction capabilities. Finally, a scale-invariant error loss is utilized to predict depth mapping in logarithmic space. Experimental results on a transmission line fault detection dataset demonstrate that the proposed method achieves the highest detection accuracy, with a mean average precision (mAP) score of 97.50% and a detection speed of up to 56 frames per second. It exhibits high robustness and accuracy, providing an effective solution for intelligent monitoring of transmission line faults.