MobileSegCut: Interactive Image Segmentation with Lightweight Networks and Fewer User Input

Image segmentation is a crucial task in computer vision but faces significant challenges related to limited computational resources and insufficient accuracy when deployed on edge devices. Although existing lightweight segmentation networks can reduce the computational burden, they often suffer from the inaccuracy due to the loss of detailed information. Meanwhile, most interactive methods rely on a large number of user inputs to improve accuracy. To address these challenges, an interactive lightweight image segmentation network MobileSegCut was proposed. It fuses deep learning with userguided optimization by combining the lightweight and efficient feature extraction capability of MobileNetV3-DeepLabV3+ with the interactive and precise segmentation capability of GrabCut. In this method, user strokes are converted into the target region information, and further integrated into the segmentation process through mask fusion and probabilistic guidance. It achieves flexible and accurate segmentation of single target and multi-target with minimal user input. Experiments conducted on the PASCAL VOC 2012 dataset show that the proposed method can accurately segment user-intended single or multiple targets with fewer than five user strokes, increasing the segmentation accuracy from 67.2% to 80.7% for the classical segmentation network.