Ke Ma, Jiaqi Tang, Bin Guo, Xueting Han, Ruonan Xu, Qingfeng He, Ziheng Wang, Xu Wang, Qifeng Chen, Zhiwen Yu, Yunhao Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

With the development of intelligent technology in coal mine industrial Internet of Things (IoT), the demand for salient object detection (SOD) in underground coal mine space has been increasing. The complex scenes and variable backgrounds in coal mine bring challenges for SOD, such as blurred edges, high computational complexity, and long processing times, making it difficult to meet the accuracy and real-time requirements of coal mine industrial IoT applications. To address these issues, we propose the dynamic weighting hybrid attention (DWHA)-partial convolution multiscale strip pooling (PCMSP) network for SOD in the coal mine industrial IoT. First, we introduce the DWHA module, which dynamically fuses self-attention for global context and SBAM for refining channel and spatial information, improving saliency detection accuracy. Second, we propose the PCMSP lightweight module, which the multiscale strip pooling introduces multiscale dilations, enhancing the ability to capture multiscale information and improve feature representation. By using partial convolution, which reduces the consumption of computing resources and running time while ensuring boundary quality. The experimental results indicate that, using self-built dataset for underground coal mine SOD, the DWHA-PCMSP network outperforms the four SOTA: BASNet, U2Net, SUCA, and EDN by achieving an increase of 2.82% in F1-score, a decrease of 23.70% in MAE, a reduction of 72.3G in FLOPs, and an improvement of 6.6 FPS in speed, compared to the worst-performing model.

Mobile Edge Computing (MEC) leverages embedded devices to deliver low-latency and cost-effective intelligent services, such as autonomous driving. Federated learning allows MEC devices to collaboratively learn a global model without exposing private data. However, the dynamic real-world environments that MEC devices operate in can lead to catastrophic forgetting of previously learned knowledge. While continual learning has been applied in federated learning to retain knowledge about past data, there are significant barriers when extending it for evolving MEC environments due to unbearable communication and computation costs for resource-constrained devices. We address this by proposing Sparse-FCL, a federated continual learning framework that uses sparse training to reduce training overhead and improve model performance. Specifically, to retain the generalization knowledge of specific tasks, we introduce a progressive neuron selection via multi-device collaboration module, which gradually selects the key neurons that show importance on multiple devices in each continual learning task. In addition, we propose a task-adaptive topology exploration module, aiming to provide more pliable dynamic sparse training configurations for the scenarios of federated continual learning. Experiments on FCL benchmarks demonstrate Sparse-FCL's superior accuracy under high sparsity levels, and it also achieves reductions of 80.4%, 89.5%, and 19.7% in communication, computation, and storage overhead, respectively, compared to existing federated continual learning methods.