Deep Learning Based Polarization Image Fusion For Metal Industrial Parts Detection

In the field of industrial manufacturing,the detection of industrial parts is not only related to economic benefits but also a matter of safety.However,defects in metal parts are often small and irregular in size,accompanied by specular reflections,making it difficult to accurately detect them with the naked eye or a regular camera.Through polarization imaging technology,it is possible to analyze the reflected light from the surface of the parts under different polarization directions,reducing the interference caused by specular reflections,thereby capturing minute surface details and improving the stability and accuracy of detection.In order to improve product quality,reduce production costs,and promote economic development,it is crucial to establish an industrial parts defect detection model based on polarization imaging.In this study,we focus on five types of defects in the collected dataset of polarization metal industrial parts.We conduct research from three aspects: polarization image demosaicing,polarization image fusion,and industrial parts defect detection..Based on deep learning,we build a series of models that enable fast and real-time detection of industrial component defects.The main content and results of this study are as follows:(1)Research on Polarization Image Demosaicing Algorithm.In this study,a color polarization camera based on a focal plane sensor is used to capture a dataset of polarization industrial parts.In order to obtain full-resolution polarization intensity component images and quantitatively describe the polarization information,demosaicing processing is required for the captured images.In this study,an improved convolutional neural network-based color polarization demosaicing model is developed by jointly applying attention mechanisms and atrous spatial pyramid pooling.Experimental results demonstrate that the improved demosaicing model achieves the highest values in both structural similarity and peak signal-to-noise ratio,while exhibiting high color accuracy.Compared to mainstream polarization demosaicing algorithms,the proposed model in this study achieves significant improvements.(2)Research on Polarization Image Fusion Model.In order to address the issues of low signal-to-noise ratio,sensitivity to noise,and the need to improve accuracy in industrial parts defect detection,this study conducts research on polarization image fusion to enhance polarization information.Considering the limited datasets and lack of ground truth fused images in polarization image fusion,this study builds a polarization image fusion model based on autoencoders.The model consists of a densely connected Convolutional Neural Network(CNN)and a Transformer model in the feature extraction layer,which captures both local and global information.It also incorporates edge detection operators to enhance the fused features.Finally,the features are fed into the decoder to generate reconstructed results.This study uses total intensity and the degree of polarization images as source images and verifies the effectiveness of the model through comparisons with 17 classical image fusion algorithms and the calculation of 9 objective evaluation metrics.Experimental results demonstrate that the polarization image fusion model achieves the best evaluations in terms of image visual quality,information richness,image contrast,and similarity to the source images.(3)Research on Metal Industrial Parts Defect Detection Model.To address the issues of scarce data samples and limited defect quantities in metal industrial parts defect detection,this study conducts data augmentation,including flipping,rotation,stretching,and other methods,to expand the dataset size.Considering the characteristics of metal industrial component defects,this study improves the YOLOv7-tiny model to maintain lightweight characteristics while enhancing detection accuracy.Specifically,this study improves the Sim-ELAN module of the backbone and applies it to the feature fusion module to enhance the representation capability of deep semantic features.Inspired by Bi FPN,this study directly introduces the features of the backbone network into the feature fusion module and designs a learnable parameter to allocate the contribution of feature maps,better fusing deep and shallow features without increasing computational complexity.Finally,considering the characteristics of polarization information,this study designs five fusion groups to verify the impact of polarization information fusion on industrial parts defect detection.Experimental results show that,without expanding the dataset,the fused polarization data can improve the average precision by up to 2.4%.A comparison with other lightweight models demonstrates that the improved industrial component defect detection model achieves a detection accuracy of up to 84% without significantly increasing the inference speed.