Accurate and quick recognition of multiple and intersecting complex power lines from images was achieved by constructing a novel encoder-decoder network. The first 16 layers of regular MobileNetV3 were taken by the encoder to reduce network parameters, and the coordinate attention mechanism was used to replace the squeeze and excitation attention mechanism of regular MobileNetV3 to obtain the channel and position information of feature maps. The multi-scale feature information of power lines was obtained by the decoder through the pyramid pooling module to improve the recognition accuracy. The encoder feature maps of 2nd, 4th, 7th, 11th and 13th layers were processed by the sharpened kernel convolution and were connected to the corresponding layers of decoder, enhancing the extraction of complex power lines edge features. The hybrid loss function was used to resolve the imbalance between classes of images with fewer power lines and more background pixels. The speed of network training was accelerated by transfer learning. Experimental results indicated that the mean pixel accuracy (MPA), mean intersection over union (MIOU) and recognition speed of the novel encoder-decoder network were 92.18%, 84.27%, and 32 frames per second, respectively, which were superior to those of PSPNet, U2Net, and other power lines recognition networks.