Abstract:

The U-Net model is large and accordingly has relatively slow speed on image processing. This drawback makes it difficult to satisfy the requirement of industrial real-time applications. On the consideration of the question above, this paper designs a light-weight full convolution neural network named LU-Net. In the proposed network, it integrates the thought of MobileNet-V2 into the U-Net framework. The depth separable convolution method in MobileNet-V2 is efficient to reduce the parameters and the computation complexity of the proposed network. The proposed network also reserves the advantages of normal convolution and bottleneck model. Accordingly, it is efficient to utilize the high-level features to keep the accuracy and reduce the processing time simultaneously. The experiments on hollow symbol dataset and DRIVE dataset indicate that, on the comparison with U-Net, the parameters of the proposed LU-Net is 0.59×106,which is 1.9% of the original model, and the processing speed is 5 times faster. Under the experimental environment, LU-Net takes only 25?ms to process a picture under the resolution of 360×270 size. LU-Net is a promising method for the industrial real-time applications on picture processing.

Key words: U-Net, semantic segmentation, MobileNet-V2, depth separable convolution

摘要：

传统U-Net网络模型大，处理图片速度慢，难以适应工业生产中实时的需求。针对该问题，设计并实现了一个轻量级全卷积语义分割网络LU-Net。LU-Net网络以U-Net框架为主体，结合MobileNet-V2的思想，利用深度可分离卷积参数少、计算量小的特点轻量化网络模型。网络综合利用bottleneck模块与普通卷积的优点，并高效利用了高层特征，在保持精度的同时，大幅缩短了分割所需时间。经公开数据集DRIVE及自制凹陷字符数据集上实验的验证，相较于原U-Net网络模型，提出的LU-Net模型参数量缩小至0.59×106，为原模型的1.9%，运行速度提高5倍，处理一张360×270图片的平均耗时为25?ms。LU-Net基本满足工业生产对图像实时处理的要求。

关键词: U-Net, 语义分割, MobileNet-V2, 深度可分离卷积