Fast Semantic Segmentation Model PULNet and Lawn Boundary Detection Method

To quickly and accurately identify the lawn area and boundary positions of different scenes, environments, and seasons, we propose a new semantic segmentation model PULNet and lawn boundary detection methods. Firstly, the ResNet50 network is improved to expand its effective receptive field, a Pooling pyramid (P) and an Upsampling dimensionality reduction structure (U) is constructed based on the Dilated_ResNet50 network. Secondly, a fast and accurate PULNet semantic segmentation network is proposed integrating the image Local detail information structure (L). Finally, an Eight-neighbor coding method is designed to accurately locate the border of the lawn. Experiments on the ADE20K dataset obtained the mean Intersection over Union (mIoU) and mean Pixel Accuracy (mPA) 32.86% and 75.65% respectively. The average speed is 82.7 frames per second on a platform with GTX 1080Ti GPU. Compared with the Fully Convolutional Network (FCN) the mIoU and mPA are increased by 3.47% and 4.33% respectively, and the speed is 11 times higher. The proposed method can be used for fast and accurate lawn semantic segmentation and boundary detection.