基于神经网络架构搜索的语义分割方法

Semantic Segmentation Method Based on Neural Architecture Search

摘要: 神经网络架构搜索旨在使用搜索策略在给定的搜索空间上让算法自动搜索出网络结构模型以减少人工设计网络的任务量，拓展神经网络架构搜索在语义分割领域的应用对自动化深度学习领域的研究有重要意义。通过设计U型搜索空间，将可微分神经网络架构搜索策略应用于语义分割模型。实验结果显示，在The Oxford-IIIT Pet数据集搜索得到的网络与基准网络UNet相比，搜索出的网络模型mIOU提高了14.1%，分割的效果更加显著，轮廓边界更加清晰。将搜索出来的网络迁移到Camvid数据集上进行测试，比基准网络实验精度提升了20.5%。研究表明，神经网络架构搜索与语义分割的结合在自动化深度学习领域的研究中具有重要意义，能够使语义分割模型获得更优秀的性能。

Abstract: The objective of Neural Architecture Search (NAS) is to use a search strategy to automatically find network structure models within a given search space, thereby reducing the task load of manually designing networks. Expanding the application of NAS in the field of semantic segmentation bears significant importance for research in automated deep learning. A U-shaped search space was de-signed, and a differentiable NAS strategy was applied to a semantic segmentation model. Experi-mental results showed that the network found on The Oxford-IIIT Pet dataset outperformed the benchmark UNet network model, with a Mean Intersection over Union (mIOU) increase of 14.1%, and produced more prominent segmentation results with clearer contour boundaries. When the discovered network was transferred to the Camvid dataset for testing, it surpassed the benchmark network experimental accuracy by 20.5%. This study demonstrated that the integration of NAS and semantic segmentation holds significant importance in the field of automated deep learning re-search. This approach enables semantic segmentation models to achieve superior performance.