ASF-Net: A Lightweight Network for Panoptic Segmentation with Adjacent Scale Feature Fusion
Chunfeng Yang, Linyuan Qi, Chenxiao Yang, Longyu Shi, Qiuyu Tang, Lixiang Tan · 2025
Panoptic segmentation is a challenging task in computer vision. However, the current convolutional neural network (CNN)-based approaches cannot effectively utilize the spatial information, while the Transformer-based approaches bring high computational complexity. To address these issues, this paper proposes an adjacent scale feature fusion module to effectively integrate adjacent scale features to enrich the information of different scales. Specifically, the features in the encoder are further extracted by a reparameterized conv block, and then weights are assigned to different features by content-guided attention. These weights are used to adjust the fusion scales to achieve more flexible fusion results. Efficient local attention is also introduced to build a lightweight network that can accurately recognize regions of interest and enhance the feature extraction capability of the backbone network. The quality of panoptic segmentation is improved by 5.9% and 2.6%, respectively, compared to Panoptic DeepLab.