Research On Serial Number Recognition For Ceramic Membranes Based On End-to-end Deep Learning

Although China has achieved great economic success,it is also a country with scarce water resources.Economic development inevitably brings waste water discharge and pollutes insufficient water resources,which not only poses a threat to people's water health,but also fails to meet the goals of sustainable development.As one of the vital components of wastewater treatment process,ceramic membrane has been widely used in various wastewater treatment scenarios.To distinguish various ceramic membranes,a uniquely identified serial number is usually engraved on the surface of each ceramic membrane.In the real wastewater treatment process,the workers install the correct ceramic membrane to the specific equipment or process according to its serial number.However,this job requires a lot of physical labor,which not only is very inefficient,but also may cause serious production accidents.In addition,the job is usually carried out in a high temperature or highly corrosive environment,making workers more vulnerable to damage to their health.In this thesis,a framework for detection and recognition of ceramic membrane serial number is proposed based on end-to-end deep learning to replace the manual job.Since there is no color difference and only depth difference between the serial number character region and the background region,a photometric stereo vision method is proposed to reconstruct the surface of ceramic membrane.First,an image acquisition system dedicated to photometric stereo reconstruction is designed to obtain ceramic membrane images irradiated from four different light sources in the same field of view,and to form the ceramic membrane image dataset.Then,the normal vector map and the divergence map of the ceramic membrane surface are acquired by photometric stereo reconstruction from 4 light sources.Experimental results indicate that the divergence map illustrates the most ideal reconstruction,which has most salient details and contours of character regions and is suitable to be used as three-dimensional(3D)shape information integrating into the end-to-end deep-learning-based framework.To suppress the inherent error accumulation of traditional step-by-step detection and recognition methods,a framework based on end-to-end deep learning is proposed for automatic detection and recognition of ceramic membrane serial number.The framework consists of three successive stages: photometric stereo reconstruction,serial number detection and recognition.Three stages are jointly trained,in which the reconstruction stage can integrate the 3D shape information of ceramic membrane serial number into the framework to improve the performance of serial number detection.To ensure the effective convergence of the framework in multi-task training,the local loss functions of three stages are theoretically analyzed,and their ranges are deduced and verified by a simulation.Experimental results show that the proposed framework achieves a better detection and recognition performance at a reasonable time consumption,with the detection index F1-score and recognition accuracy of95.61% and 96.49%,respectively.The proposed end-to-end deep-learning-based framework for ceramic membrane serial numbers can play a certain role in promoting the automation and intelligence of ceramic membrane production management and wastewater treatment process.