【摘要】 近年来，基于二阶段模型的图像分割网络凭借其卓越的性能和鲁棒性，在自动驾驶、医疗等多个领域得到广泛应用.然而，这类网络的庞大模型和复杂计算量严重限制了它们在低算力、低功耗的移动端嵌入式平台上的部署.为此，本文提出了一种新的轻量级二阶段分割网络CDViT Mask R-CNN,通过采用廉价卷积和长距离解耦注意力机制(DFC)对Mask R-CNN这一图像分割领域主流模型的backbone进行重构，在平均精度(mAP)仅下降了0.4%情况下，模型整体尺寸缩减了46.5%,推理帧率(FPS)提升了12.6%.此外，本文借助DFC特性，采用基于掩模恢复的知识蒸馏策略对模型进行多尺度蒸馏，以补偿模型轻量化后的精度损失，使蒸馏后的模型精度提高了1.2%.实验结果表明，本文提出的模型在分割任务中具有更好的速度与精度权衡.更多还原

【Abstract】 In recent years, image segmentation networks based on two-stage models have been widely applied in various fields, such as autonomous driving and healthcare, due to their outstanding performance and robustness.However, such networks’ large models and complex computational requirements severely limit their deployment on low-power, low-computing platforms, especially mobile embedded systems.Therefore, we propose a novel lightweight two-stage segmentation network called CDViT Mask R-CNN.Adopting cheap convolution and a long-range decoupled attention mechanism(DFC) to reconstruct the backbone of the Mask R-CNN reduces the overall model size by 46.5% and improves the Frames Per Second(FPS) by 12.6%,while the average precision(mAP) only decreases by 0.4%.Furthermore, leveraging the DFC characteristics, a multi-scale distillation strategy based on mask recovery is employed to compensate for the accuracy loss caused by model lightweighting, resulting in a 1.2% improvement in the distilled model’s precision.Experimental results demonstrate that the proposed model achieves a better trade-off between speed and accuracy in segmentation tasks.更多还原