Lightweight Real-Time Polyp Segmentation Network Based on Factorized Convolution Unit

Abstract

The current polyp segmentation network generally uses ResNet and Res2Net as the backbone extraction network, which improves segmentation accuracy. However, due to its high computational burden and hundreds or thousands of feature channels, it limits the deployment of the network on devices with limited computing power, such as embedded devices and mobile devices. Based on the above requirements, a lightweight real-time polyp segmentation network is proposed, which uses factorized convolution unit as the basic extraction unit of the backbone network, which dramatically reduces the computational complexity and the number of parameters while maintaining the segmentation accuracy. Specifically, the proposed model has only 679,000 parameters and runs at 73FPS in a single RTX 3060 GPU to achieve real-time segmentation. Experiments show that the method in this paper achieves the trade-off of speed, network size, and accuracy in Kvasir-SEG and ClinicDB datasets, with IoU reaching 81.72% and 80.41%, respectively