Research On Classification Method Of Micro-Cracks Images Of Sheet Metal Bending Parts

Sheet metal bending as an important process method,has the characteristics of rapid prototyping and simple process.Inthe process of sheet metal bending,the external surface of the corner is subjected to tensile force and easily generates micro-cracks,even large cracks.The quality of the sheet metal bending parts at the corners directly affects the product's stability and service life.Therefore,the classification of micro-cracks at the corners of sheet metal bending parts has great significance for the improvement of the process and the quality assurance of the sheet metal bending products.At present,the quality inspection of the sheet metal bending part mainly uses the shape measurement equipment such as a laser radar to obtain the outer dimensions of the sheet metal part and determine whether the surface has a defect by artificial visual inspection method.Artificial visual inspection methods are not only inefficient,but also subject to subjective judgment.This paper studies the classification method that using deep learning for classifying micro-cracks at the bend angles of sheet metal bending parts.The main research work includes:Designed and built an image acquisition system for the bend angles of sheet metal bending parts.The images at sheet metal bending parts corners were divided to 5 classes according to ASTM B820-98,and divided dataset into training set and testing set randomly.Using the traditional machine vision method to extract the features of the defect pictures,analyze the feature different among the different defect pictures,and use the gray feature and texture analysis methods to classify pictures of different grades of defects.A method of defect classification based on deep learning was proposed.Designed a convolution neural network model to classify images with different image preprocessing methods and different size of image and different data augmentation.The results were obtained through compared experiments.A method of transfer learning based on Inception-V3 was proposed to classify sample sets.The pre-trained Inception-V3 model is used as a feature extractor to obtain feature vectors that have 2048-dimensional.Compared three different classifiers to classify the feature vectors.The results showed that the method of transfer learning combine with MLP got 98.18%accuracy that was the best result.