Research On 3D Quantitative Defects Recongnition Method In Continuous Slab Surface Based On Deep Convolution Neural Network

Continuous casting,as the main process of liquid steel solidification in steelmaking process,has been widely promoted and applied.As the intermediate semi-finished product of deep processing of steel materials,the surface or internal quality of continuous casting billet directly determines the performance and market value of the final product,and improves the production cost of enterprises.Therefore,for the unsteady production process,intelligent product quality online detection method can effectively improve the production efficiency and energy saving level of enterprises.Avoid low efficiency,poor real-time performance,false detection and missed detection caused by manual detection.In this thesis,aiming at semi-finished steel material(continuous casting billet),with intelligent online detection as the research goal,a real-time detection method of defect 3D morphology based on binocular CCD laser scanning imaging and deep learning fusion is proposed.The main research contents are as follows:1.According to the requirements of product detection,a set of binocular CCD laser scanning imaging test platform is designed and developed,including the design of system imaging structure and 3D imaging algorithm.In this platform,the laser monoline triangulation is used as the depth measurement method,and the information extraction method of low occlusion rate defects is constructed through the binocular image fusion module,which provides a platform support for the end-to-end training and recognition model construction of deep neural network based on 3D morphology defect detection.2.According to the three-dimensional scanning image of the system,a fast locating method of suspicious region(Region Of Interest,ROI)based on three-dimensional depth information of image surface is proposed.Among them,the generation of defect candidate boxes to be identified based on the initial location of ROI region information can reduce the calculation amount of the traditional deep learning candidate box generation,and solve the problem that small samples cannot be trained,so as to support the effective training and recognition of the local characteristics of the defect region,eliminate the pseudo-defect interference on the product surface to the maximum extent,and improve the real-time and accuracy of defect detection.3.According to the imaging and detection characteristics of the system,a deep learning defect recognition model of full convolution and full connection fusion is proposed.According to the initial location of the defect,a two-step defect recognition and segmentation model is adopted.That is,the fully connected network model is used to identify and classify the initial location of the candidate frame area,and then the recognition result is sent to the fully convolutional network for regional semantic segmentation,which improves the accuracy of defect quantitative detection.4.The feasibility of the system design method is verified by experiments.The system adopts end-to-end training method,and trains and tests with unified data representation.The DICE coefficient of system defect location recognition and segmentation can reach more than 93 %.Ge Force GTX 1080 GPU can perform 15 image recognition and segmentation tasks per second.