Abstract: The advancement of deep learning has boosted the research on image semantic segmentation. At present, most effective methods for this research are based on the fully convolutional neural networks. Although the existing semantic segmentation methods can effectively segment the image as a whole, they cannot clearly identify the edge information of the overlapped objects in the image, and cannot effectively fuse the high- and low-layer feature information of the image. To address the above problems, superpixel segmentation was employed as an auxiliary optimization to optimize the segmentation results of object edges based on the fully convolutional neural network. At the same time, the design of a joint cross-stage partial multiscale feature fusion module can enable the utilization of image spatial information. In addition, a skip structure was added to the upsampling module to enhance the learning ability of the network, and two loss functions were adopted to ensure network convergence and improve network performance. The network was trained and tested on the public datasets PASCAL VOC 2012. Compared with other image semantic segmentation methods, the proposed network can improve the accuracies in pixel and segmentation, and displays strong robustness. 

Key words:  fully convolutional neural network, multiscale feature fusion, superpixel segmentation