The traditional vision algorithms have low accuracy and are greatly affected by environmental factors during the detection of long and slender power lines in complex scenes, and the existing power line detection algorithms based on deep learning are not efficient. In order to solve the problems, an end-to-end fully convolutional neural network model was proposed which was suitable for power line detection on mobile terminals. Firstly, a symmetrical encoder-decoder structure was adopted. In the encoder part, the max-pooling layer was used for down-sampling, so as to extract multi-scale features. In the decoder part, the max-pooling indices based non-linear up-sampling was used to fuse multi-scale features layer by layer to restore the image details. Then, a weighted loss function was adopted to train the model, thereby solving the imbalance problem between power line pixels and background pixels. Finally, a power line dataset with complex background and pixel-level labels was constructed to train and evaluate the model, and a public power line dataset was relabeled as a different source test set. Compared with a model named Dilated ConvNet for power line semantic segmentation on mobile devices, the proposed model has the prediction speed for 512×512 resolution images on the mobile device GPU NVIDIA JetsonTX2 twice that of Dilated ConvNet, which is 8.2 frame/s; the proposed model achieves a mean Intersection over Union (mIoU) of 0.857 3, F1 score of 0.844 7, Average Precision (AP) of 0.927 9 on the same source test set, which are increased by 0.011, 0.014 and 0.008 respectively; and the proposed model achieves mIoU of 0.724 4, F1 score of 0.634 1, AP of 0.664 4 on the public test set, which are increased by 0.004, 0.007 and 0.032 respectively. Experimental results show that the proposed model has better performance of real-time power line segmentation on mobile terminals.