Cloud-edge efficient image segmentation via lightweight multiscale attention learning

Image segmentation in cloud-edge visual analytics needs to balance pixel-level accuracy with practical constraints such as inference latency, model size, and communication cost. Most existing segmentation models tend to rely on heavy backbone networks or complicated feature fusion designs to improve accuracy, which limits their applicability on edge devices with restricted computing resources. To alleviate this problem, this paper presents LMANet, a lightweight multiscale attention network for efficient image segmentation. In LMANet, an efficient convolution-based compact encoder is used for feature extraction, and a multiscale attention module is further designed to strengthen boundary details, regional contextual information, and global semantic responses. With its encoder-attention-decoder architecture, basic feature extraction can be performed on the edge side, while enhanced feature aggregation and decoding can be flexibly supported by the cloud side when needed. Experiments on the ISIC 2018 dataset demonstrate that LMANet achieves a Dice score of 0.921 and an IoU of 0.853, with only 5.6 M parameters and 9.4G FLOPs. Compared with representative segmentation methods, LMANet shows a more balanced performance in terms of segmentation accuracy, model compactness, and inference efficiency. These results demonstrate the effectiveness of LMANet for computationally efficient skin-lesion segmentation and indicate its potential for resource-constrained cloud-edge visual analytics.