Abstract: In industrial surface QC scenarios, deep neural networks classify product images for qualified judgment or quality grading. Surface QC equipment equipped with deep classification neural networks is required to carry out Attribute Reproducibility and Repeatability (AR&R) assessment of gauges, but the deep classification neural network won't be able to output consistent classification results and blurring results for disturbed images.Due to assembly tolerance, equipment vibration, and other factors, captured product images exhibit perturbations in position, angle, brightness, and blurring. Consequently, the classification neural network cannot consistently output classification results and probabilities for these perturbed images, which results in the surface QC equipment failing to pass the AR&R assessment, which is summarised as a network output reproducibility problem, we propose a Siamese network-based classification neural network training method, the Siamese primary network uses the original samples for supervised learning training to learn to output the correct classification categories, the Siamese secondary network copies the primary network weights through exponential smoothing and outputs feature embeddings of the perturbed samples corresponding to the original, which is used for comparative learning training of the primary network, so that the primary network outputs a consistent classification probability for both the original samples and the perturbed samples inputs, When reasoning, only the main network is retained for product defect classification. In order to fully verify the performance of the algorithm, benchmark experiments, network architecture ablation experiments and comparison experiments with similar methods are designed and verified on inductive product images, and the classification accuracy reaches 99.3462% with a classification probability variance of 0.001016 in the verification results, The described method effectively addresses the output repetitiveness issue of deep classification neural networks for industrial product image classification, reducing classification probability variance and improving accuracy.

摘要： 工业表面质检场景中，深度分类神经网络常用于对产品图像进行分类，实现对产品的合格判别或品质分级，搭载深度分类神经网络的表面质检设备需进行量具的检验重复性与再现性评估（Attribute Reproducibility and Repeatability，AR&R），但由于产品载具存在装配公差以及设备振动等因素，设备拍摄的产品图像会出现位置、角度、亮度、模糊度的相关扰动，分类神经网络对扰动图像将无法输出一致的分类结果和分类概率，导致表面质检设备无法通过AR&R评估，将此问题总结为网络输出重复性问题，并针对这类问题提出了一种基于孪生网络的分类神经网络训练方法，孪生主网络使用原始样本进行监督学习训练，学习输出正确的分类类别，孪生次网络通过指数平滑拷贝主网络权重，输出与原始对应的扰动样本的特征嵌入，用于对主网络进行对比学习训练，使主网络对原始样本与扰动样本输入均输出一致的分类概率，推理时，仅保留主网络用于产品缺陷分类。为充分验证算法性能，设计了基准实验、网络架构消融实验与类似方法对比实验，并在电感产品图像上进行了验证，验证结果中分类准确率达到99.3462%，分类概率方差为0.001016，所述方法可有效缓解使用深度分类神经网络对工业产品图像分类的输出重复性问题，在显著降低了分类概率方差的同时，分类准确率也有一定提升。