一种基于选权迭代的样本数据自动清洗方法

An Automatic Sample Data Cleaning Method Based on Weight Iteration

基于数字线划图和真正射影像可以自动生成大量满足深度学习要求的样本数据,但是其中存在部分错误信息,会导致神经网络模型训练难度加大,并且限制地物提取精度提升。因此,提出一种基于选权迭代的样本数据自动清洗方法,首先构建数据清洗深度神经网络模型,并给出基于选权迭代的网络训练方法,该方法打破传统网络模型训练时认为所有样本对损失函数计算权重相同的假设,利用数据清洗网络模型训练过程中样本预测精度作为样本权重并带入网络训练中,再通过迭代训练不断更新样本权重,最终将权重低的样本剔除,以实现自动数据清洗和样本库精化。使用数据清洗前后的样本库,对5种经典语义分割网络模型进行训练和精度对比实验,结果表明,利用数据清洗后的样本库训练模型,建筑提取精度平均提高2.36%,道路提取精度平均提高3.48%,水体提取精度平均提高1.88%,证明提出的数据清洗方法可以有效提高网络模型的提取精度。

Based on digital line graphic and true digital orthophoto map,a large amount of sample data that meets the requirements of deep learning can be automatically generated.However,there are often some erroneous information,which increases the difficulty of training neural network models and limits the improvement of ground feature extraction accuracy.A sample data automatic cleaning method based on selection weight iteration was proposed to address this issue.Firstly,a deep neural network model for data cleaning was constructed,and a network training method based on selection weight iteration was proposed.The method broke the assumption that all samples had the same weight for the calculation of loss function during network model training.The prediction accuracy of the samples during the data cleaning network model training process was used as the weight of the samples to be brought into the network training,and the sample weights were continuously updated through iterative training.Finally,samples with low weights were eliminated to achieve automatic data cleaning and sample database refinement.Training and accuracy comparison experiments were conducted on five classic semantic segmentation network models using a sample database before and after data cleaning.The results show that,the model trained using the sample database after data cleaning improves the average accuracy of building extraction by 2.36%,road extraction by 3.48%,and water extraction by 1.88%.This experiment proves that the data cleaning method proposed in this paper can effectively improve the accuracy of the network model in extracting ground features.