The ceramic membrane plays an important role in the wastewater disposal industry. The serial number engraved on each ceramic membrane is an essential feature for identification. Here, an automatic inspection system for serial numbers of ceramic membranes is proposed to replace the manual inspection. To the best of our knowledge, this is the first attempt to automatically inspect serial numbers of ceramic membranes. To suppress error accumulation inherently existed in the previous stepwise approaches, an end-to-end photometric-induced convolutional neural network framework is proposed for this automatic inspection system. The framework consists of three sequential stages, which are photometric stage for performing photometric stereo, localization stage for localizing the text region, and recognition stage for producing recognition results. The photometric stage can integrate three-dimensional shape information of serial numbers of ceramic membranes into the framework to improve the inspection performance. Since three stages are jointly trained, a theoretical analysis on the contributions of the local losses is provided to ensure the convergence of the framework, which can guide the design of the total loss function of the framework. Experimental results demonstrate that the proposed framework achieves better inspection performance with a reasonable inspection time compared with the state-of-the-art deep learning methods, whose localization performance and recognition performance are the F-score of 95.61% and the accuracy of 96.49%, respectively. Furthermore, these demonstrate the potential that our proposed automatic inspection system will be beneficial for the intelligentialize of the ceramic membrane manufacturing and wastewater treatment if it is equipped with a perception system and a control system in ceramic membrane production lines and wastewater treatment processes.

陶瓷膜在废水处理行业中起着重要作用。刻在每个陶瓷膜上的序列号是识别的基本特征。在此，提出了一种用于陶瓷膜序列号的自动检查系统来代替手动检查。据我们所知，这是自动检查陶瓷膜序列号的首次尝试。为了抑制先前逐步方法固有的错误累积，针对此自动检查系统提出了端到端的光度诱导卷积神经网络框架。该框架包括三个连续的阶段，分别是用于执行光度立体的光度阶段，用于对文本区域进行本地化的定位阶段以及用于产生识别结果的识别阶段。光度平台可以将陶瓷膜序列号的三维形状信息整合到框架中，以提高检测性能。由于三个阶段是共同训练的，因此对局部损失的贡献进行了理论分析，以确保框架的收敛性，从而可以指导框架总损失函数的设计。实验结果表明，与最新的深度学习方法相比，该框架的定位性能和识别性能为F.评分为95.61％，准确性为96.49％ ， 分别。此外，