Since the results of traditional Convolutional Neural Network on account of the image segmentation tasks is not satisfying enough, a Splicing Pyramid Scene Parsing Network (SPSP Net for short) is put forward based on PSP Net and U-Net. Through multiple splicing in the process of up-sampling, the loss of image information can be effectively reduced during information coding. At the same time, multiple feature fusion is adopted to improve the expression of effective image information. Finally, SPSP Net is simulated and verified on the Zebra crossing dataset. The experimental results show that the proposed SPSP Net effectively improves the accuracy of image semantic segmentation compared with U-Net, Seg Net, and PSP Net, and the segmentation effect for the image edge part is also significantly enhanced.