Research On Object Detection And Classification Algorithms For Surface Defects Of Strip Steel

Steel plate has a very wide range of applications in our lives,and it is widely used in some chemical equipment,light industry manufacturing,aerospace and other engineering fields.In industrial production,due to the irresistible factors of material properties and processing technology,there will be different types of defects on the surface of strip steel,such as rolling scale,emulsion spots,cracks,pits,inclusions,scratches and so on.These defects not only affect the quality of products,but also affect the application of products.The surface defects of plate and strip steel are one of the main reasons for the poor processing quality of plate and strip steel.The defect of the surface of the strip steel can not only make the corresponding products become inferior products,but also cause serious damage to the roller,and even more serious will cause production accidents,which will have a very negative impact on the production of the corresponding enterprises due to the surface damage of the strip steel,thus causing huge economic losses.It may also have many adverse effects on society.Therefore,the efficient and accurate detection of strip surface defects is the key to improve the quality of strip products,and the detection method has more important research value.In this thesis,the classification and detection algorithms of strip steel surface defects are studied,and an improved limit learning machine and a strip steel defect detection network model based on semantic segmentation and limit learning machine are proposed.The main research contents and results are as follows:(1)Aiming at the problem of unstable results caused by random parameters,this thesis uses particle swarm optimization(PSO)to optimize the parameters of the extreme learning machine(ELM).At the same time,due to the different reasons for the formation of strip surface defects,strip steel defects are presented in various forms,and have the characteristics of randomness and complexity.To solve this problem,this thesis uses ReliefF combined with PCA algorithm to screen and reduce dimensions of features,remove redundancy and improve classification accuracy.Experiments show that the recognition rate of the six types of strip steel surface defects using this algorithm reaches 97.2%,which is 6% higher than the original ELM algorithm alone.At the same time,the speed of this algorithm can also meet the online detection requirements of surface defects.(2)In this thesis,the relevant research,the obtained image information through the corresponding technology into data information,in order to follow-up experiments,uses the LBP algorithm to extract the defect features,and then saves these feature data into a csv file and manually adds tags to make a new data set.In addition,in the absence of a publicly available data set with real images of labeled surface defects,the real strip steel surface defect images were labeled using the Labelme tool.Constructed a strip steel surface defect detection data set based on semantic segmentation and extreme learning machine.(3)Aiming at the problem that the surface defect characteristics of strip steel are not obvious,the end-to-end target detection deep learning method is introduced into the surface defect detection.When applying the deep learning method,tens of thousands of samples are needed,but in reality,it is often very difficult.Hard to get.In response to the above problems,this thesis proposes a network model based on semantic segmentation and extreme learning machine(ELM)to achieve end-to-end target detection of strip surface defects.In the experiment,the test accuracy rate reached 98.5%,which fully met the requirements of strip steel surface defect detection.