利用通道剪枝技术的实时实例分割方法

随着实例分割技术在各种场景中的应用越来越广泛，运行速度和硬件资源占用是该技术在应用中需要考虑的2个重要因素。最近提出的基于图像原型掩码系数的实例分割网络(YOLACT)在运行速度方面做得很好，但是需要设置较大的特征提取网络才能保证分割精确度，这就导致了模型占用的硬件资源较多，同时运行速度也受到了限制。在YOLACT的基础上，提出一种新的模型，对实例分割的特征提取网络进行了优化，先使用基于批量归一化层放缩因子的通道剪枝方法对YOLACT网络进行压缩，然后对压缩后的卷积层和批量归一化层进行融合，最后，在COCO val2017上对本文提出的方法进行了评估。实验结果表明，相比原始的YOLACT网络，该方法的模型文件大小可以减少56.9%，运行速度提升28.6%，运行时显存占用也降低了13.6%，有效地减少了硬件资源占用，并且提升了运行速度。

Real-time instance segmentation using channel pruning

With the application of instance-segmentation in various scenarios, the running speed and the utilization of hardware resources are two important factors to be considered in the application of instance segmentation. Recently, a instance segmentation network named You Only Look At Coefficients(YOLACT) bears a high processing speed. However, YOLACT needs to set a large feature extraction network to ensure the segmentation accuracy, which leads to high resource occupancy and limited running speed. Based on the YOLACT, a new model is proposed, and the feature extraction of network segmentation is optimized. Firstly, a channel pruning method based on batch normalized scale factor is utilized to compress YOLACT network, then the convolution layer and batch normalization layer are fused. Finally, the proposed approach is evaluated on Common Objects in Context(COCO) val2017. The experimental results show that, the model size of the method can be reduced by 56.9% and the running speed can be improved by 28.6% compared with that of the original YOLACT network. This approach can reduce the hardware resource consumption and can improve the running speed.