Analysis and Optimization of CNN-based Semantic Segmentation of Satellite Images
Heeji Im, Hoeseok Yang · 2019
In this paper, we analyze and optimize the CNN (Convolutional Neural Network) based semantic segmentation network, U-Net. While U-Net is known to be the state-of-the-art in semantic segmentation, it is difficult to apply U-Net directly to the resource-constrained systems due to its high memory usage and computation requirement. To overcome this disadvantage, we apply the filter pruning method to optimize the network. Experimental results show that the memory usage is reduced by 0.26 times and the inference speed is enhanced by 0.57 times with 75% filter pruning. In addition, the IOU (Intersection Over Union) and F1-score which reflect the accuracy of semantic segmentation, were only reduced by 4.7% and 4.5% respectively compared to the original U-net.