基于模型剪枝的轻量化人像抠图研究

Research on Lightweight Portrait Matting Based on Model Pruning

近年来,人像抠图在计算机视觉领域取得了许多进展。作为底层视觉任务,人像抠图具有较高计算量,导致模型难以部署在计算资源有限的设备上。针对该问题,文章提出自适应和固定比例相结合的剪枝方法,对人像抠图网络MODNet进行压缩。实验表明,该方法可以基本保持模型精度,将MODNet参数量和计算量分别降低了79.22%和76.09%,在中央处理器(Central Processing Unit,CPU)和NSC2上的推理速度分别提高了93%和65%。

In recent years,portrait matting has made many advances in the field of computer vision.As a low-level visual task,portrait matting has a high computation cost,which makes it difficult to deploy models on devices with limited computing resources.To address this problem,this paper proposes a pruning method combining adaptive and fixed-ratio to achieve compression of MODNet.Experiments show that the method reduces the parameter and computation cost of MODNet by 79.22%and 76.09%,while maintaining the accuracy of the model.Additionally,the inference speed on Central Processing Unit(CPU)and NSC2 increased by 93%and 65%.