| Optical filters are an important component of optical instruments,used to select specific radiation bands and are widely used in highly sophisticated products.In order to ensure the normal use of filters,certain requirements are placed on their surface quality.At present,the detection of surface defects on optical filters is mainly manual,which has problems such as the influence of subjective factors,inconsistent detection standards and low detection efficiency.Relying solely on manual inspection,the effectiveness and efficiency of inspection is no longer adequate for current production needs,so work to improve the efficiency of filter defect detection is essential.The main difficulties in the detection of defects on the surface of optical filters are the variety of product forms and sizes,and the many types of defects with small targets.In response to these problems,this paper analyses the requirements for optical filter surface defect detection and proposes a filter surface defect detection and pinpointing algorithm to improve the efficiency and accuracy of optical filter surface quality inspection in practical applications.The main work is as follows:(1)Based on the built hardware platform for optical filter surface defect detection,the motion control module and the image acquisition module of the surface defect detection system are designed.The characteristics of the filter surface defects are analysed,and suitable industrial cameras,lenses and lighting solutions are selected to achieve batch acquisition of optical filter defect images and automatic screening of inferior products.(2)A deep learning based optical filter surface defect detection algorithm is designed.Firstly,based on the image acquisition module designed in this paper,the batch acquisition of defective images of optical filters is completed,while the annotation of defective images is completed using the annotation tool,and the dataset of optical filter surface defects is constructed.Secondly,the HALCON semantic segmentation model and the U-net network model are introduced,and the defect dataset constructed in this paper is used to train the depth model and to analyse and compare the performance of the model and the shortcomings of the application to optical filter defect detection.Finally,the improved methods of multi-scale input,attention mechanism and multi-scale fusion are proposed to address the shortcomings of the network model and the characteristics of small target and large scale difference of filter surface defects.The experimental results show that the improved model network improves the segmentation accuracy of small target defects by 9.9%.(3)An optical filter surface defect pinpointing algorithm is proposed.Filters need to be priced and sold according to surface quality grades in actual production,and their quality grading is determined mainly based on the number and location of defects.A defect location algorithm has been designed for different filter shapes due to hardware errors and camera limitations in the batch inspection of filters.The results of the filter defect detection algorithm and the precise location of the defects are used to determine the surface quality of the filters.Comparison experiments with manual inspection show that the detection speed of the algorithm proposed in this paper is 5 times faster than manual inspection,and the correct detection rate is95.7%,which can meet the needs of actual industrial production.(4)The software for the optical filter surface defect detection system was developed.The requirements of the optical filter surface defect detection system were analysed,and the design of the system software functions was carried out on the basis of the requirements analysis.The integration of defect detection algorithms and defect pinpointing algorithms was completed on the defect detection system equipment,and the software system was developed.A simple and friendly human-computer interaction interface was also designed to visualise the inspection results. |
PDF Full Text Request