Research On Surface Defects Recognition Method Of Strip Steel Based On Deep Learning

After the stage of incremental expansion focusing on capacity,the steel industry in our country has entered an optimization phase,paying more attention to quality rather than output.They are increasingly concerned with improving the industrial structure,upgrading equipment and improving product quality.As the surface quality of strip steel is directly related to the product quality,it is of great significance to develop a more efficient,more intelligent and more accurate surface defect recognition system.Steel enterprises not only need to meet customers’ constantly improving quality requirements,but also need to continuously improve the production efficiency to adapt to the fast-paced market.Therefore,they urgently need more accurate and faster defect recognition algorithm.In this paper,the defect recognition algorithm is deeply studied,which has higher efficiency,higher precision and stronger adaptability under complex background.For the characteristics of strip steel defects,a surface defect recognition scheme based on deep learning is proposed An on-line recognition platform for strip surface defects was designed and a supporting defect recognition software platform is developed.The main research results of this paper are as follows:(1)A strip surface defect recognition algorithm based on deep learning and multi-scale feature fusion was proposed.Under the condition of small sample and strong interference,a self-learning model of defect features was established,which solves the detection problems caused by the variety of surface defects,variable morphology,and small size of strip steel.The experimental results show that the algorithm can identify the defects of strip steel well,especially in the recognition of scratches,inclusions and other small and slender defects.(2)Aiming at the problem of small sample and inconsistent sample distribution,the dynamic data amplification method is adopted to simulate the real scene.The multi-task loss function of defect identification is defined and the validity of transfer learning is confirmed through experiments.As the optimizers were studied in detail the advantages and disadvantages of different optimizers are compared.During the training,in order to improve the training efficiency,the model parameters are initialized by pre-training and fine-tuning is adopted.The Mini-Batch gradient descent method is selected to obtain the global optimal solution and the learning rate attenuation method is adopted to accelerate the training(3)An online recognition system for strip surface defects was designed,including a mechanical motion system and a visual system.Among them,PLC is the main control to connect the servo motor and servo controller in the mechanical motion system.The light sources,cameras and lens are selected as the vision system required.High quality strip image acquisition in moving state is realized,based on the vision system.Besides,a software platform for defect recognition is developed,which includes image acquisition,defect recognition and defect analysis.(4)In order to verify the effectiveness of the model in a real scene,an online test was conducted to simulate the actual operating conditions.The strip defect samples were collected by machine vision system.After deploying the freeze model to the system,mAP and FPS indicators are used to evaluate the recognition accuracy and speed of the system.The mAP of strip surface defect detection is 98%,and the detection speed reached 18FPS,which can meets the requirement of online recognition for strip steel defects.