S2DiNet: Towards lightweight and fast high-resolution dichotomous image segmentation

Abstract

The Dichotomous Image Segmentation task aims to achieve ultra-high precision binary segmentation for category-agnostic objects, including salient, camouflaged, structurally complex, or feature-similar entities. Traditional methods designed for low-resolution inputs produce blurred segmentation, failing to meet such critical safety and stability requirements. Although existing DIS methods achieve high accuracy, they are often parameter-heavy and slow, neglecting practical application needs. To address these challenges, this paper proposes a light-weight and fast framework, aims at improving processing efficiency while ensuring accuracy in high-resolution natural scenes. The proposed method utilizes a shared-weight ResNet-18 backbone to process inputs of different scales. A Feature Synchronization module is employed to enhance the correlation between encoded features of different resolutions. To reduce the parameter and increase the inference speed, the number of feature channels are decreased; however, this also resulted in information loss. The Star Fusion module is introduced to mitigate this issue. Furthermore, a Decoupling and Integration Decoder is adopted to progressively decode and fuse the body, detail, and mask features of the object, enhancing feature decoding accuracy. The proposed model runs at 26.3 FPS with a 48.7 MB size, reducing parameters by 72.4% and increasing speed by 30.8% compared to baseline method ISNet, while maintaining superior performance. Moreover, it surpasses several existing high-resolution methods in terms of accuracy.