Research On Surface Defect Detection Of Continuous Casting Slab Based On Machine Vision

In the process of continuous casting billet production,it will produce cracks and other defects on the surface of continuous casting billets due to technological reasons,which will produce steels with quality defects when it enters the hot rolling process.This will bring economic losses to enterprises for the steels with quality defects cannot be sold.At present,the surface defect detection of continuous casting billet mostly depends on manual visual inspection,which only uses human eyes to detect slab surface defects,with high labor intensity and harsh working environment.It is easy to cause fatigue,miss inspection and false inspection when human eyes detect slab defects for a long time.Moreover,the manual visual inspection method is only suitable for offline inspection,which cannot meet the high-speed production rhythm on the production line.Therefore,a detection method based on machine vision and YOLOv3 target detection technology is proposed aiming at the real-time surface defect detection of continuous casting billets.The main research contents are as follows:1)Acquisition of surface defect image of continuous casting billet.In order to realize the real-time detection of surface defects in the process of continuous casting billet production,a machine vision acquisition system is built combining the actual production environment,the camera,lens,light source,filter,acquisition control equipment and algorithm server.Besides,a software program for surface inspection of continuous casting billets is designed and developed by analyzing the defect morphology of the collected images.2)Image processing.In order to preprocess the collected images,different smoothing filters and image enhancement algorithms for the noise and defects in the images are studied and compared firstly.Then,the preprocessed images are compared and the algorithm with obvious effect is selected to meet the preprocessing of the image.3)The surface defect detection technology of continuous casting billet based on deep learning is studied.Adopting YOLOv3 network model,this paper not only introduces the convolution neural network structure,classical convolution neural network model and target detection algorithm in deep learning theory,but also analyzes the algorithm principle of YOLOv3 in detail.Then,the YOLOv3 network model is trained using the training set.After training,the detection accuracy index mAP of the network model reaches as high as 97.30%,with a detection speed of 47 FPS.4)An improved surface defect detection method for YOLOv3 continuous casting billet is proposed.In view of the fact that the detection speed of the original YOLOv3 can not meet the needs,the algorithm is improved lightweight using,MobileNetv2 as the feature extraction network to optimize the network structure and reduce the parameters,introducing cooperative attention mechanism to strengthen the feature extraction of network and taking CIo U loss function instead of IoU bounding box regression loss.The improved algorithm achieves 96.96% mAP and 91 FPS detection speed.The weight obtained by model training is only 30 M,which meets the real-time detection requirements of continuous casting billet.Figure 29;Table 14;Reference 55...