基于特征过滤与特征解耦的域泛化模型

Feature filtering and feature decoupling based domain generalization model

针对跨场景情况下图像亮度不一致导致的深度缺陷检测模型泛化性能差的问题,提出基于特征过滤与特征解耦的域泛化(FF-FDDG)模型.模型包含设计的亮度过滤-残差模块(LFR),该模块通过实例归一化过滤亮度变化特征,并从被过滤的特征中提取缺陷高相关性且亮度低关联性的特征,并将这些特征与实例级归一化后的特征进行融合,以增强模型在跨场景图像亮度变换情况下的泛化能力.提出对比白化损失(CWL)函数,该函数通过解耦特征中亮度变换特征和缺陷纹理特征,引导模型学习缺陷纹理特征,以提升模型泛化能力.在从光伏电池制造环境中收集的跨场景光伏电池表面缺陷数据上进行实验,结果表明,相较于现阶段最先进的域泛化模型,所提出的FF-FDDG在跨场景情况下的平均检测精度(mAP)均值提升5.01%.

A feature filtering and feature decoupling based domain generalization model(FF-FDDG)was proposed,aiming at the problem of poor generalization performance of the deep defect detection model caused by inconsistent image brightness across scenes.A designed luminance filtering-residual module(LFR)was included in the proposed model.The brightness variation features were filtered out through instance normalization,the features with high defect correlation and low brightness correlation were extracted from the filtered features,and these features were fused to enhance the generalization ability of the model under the condition of cross-scenario image brightness transformation.Furthermore,a contrast whitening loss(CWL)function was proposed,by which the model was guided to learn the defect texture feature by decoupling the brightness transform feature and the defect texture feature,so as to improve the model generalization ability.The experimental results on the cross-scenario surface defect data collected in the photovoltaic cell manufacturing environment showed that,compared with the state-ofthe-art domain generalization model,the average mean average precision(mAP)of the proposed FF-FDDG model in cross-scenario situations was improved by 5.01%.