面向语义分割机器视觉的AutoML方法

AutoML method for semantic segmentation of machine vision

自动机器学习(Automatic Machine Learning,AutoML)可实现语义分割,使机器学习大部分步骤自动化。针对面向超参数优化、迁移学习、神经架构搜索等方法的算法思想、优化对象、实现技术、技术指标、应用效果及场景,结合语义分割的机器学习超参数多、数据集规模较小、标注工作量大等问题,指出超参数优化、迁移学习、神经架构搜索分别有助于提升训练效率、降低样本标注工作量、自动构建专用卷积神经网络,若Au-toML与机器视觉相结合可赋予系统自学习、快速更换检测对象和解决特别复杂任务等特性。

Automatic machine learning(AutoML)can automate most machine learning process of semantic segmentation.This paper discusses hyper-parameter optimization,transfer learning,and neural architecture search methods.The differences among algorithm thought,optimization objects,deployment methodology,technical indicators,application effects and scenarios of these methods are compared.The problems of the machine learning of semantic segmentation have been pointed out,such as multiple super parameters,small data set and large labeling workload.Draw the conclusion that the super-parameter optimization,transfer learning and neural architecture search are helpful to improve training efficiency,reduce the workload of sample labeling and automatically construct special convolutional neural network,respectively.If combined AutoML with machine vision,it can endow the system with characteristics such as self-learning,rapid replacement of test objects and solving complex tasks.