A Research Of Defect Detection On The Surface Of Parts Based On Computer Vision And Its Application

With the continuous development of computer vision and machine learning,some-thing has been changed.In the fields of computer vision and machine learning,object detection is a very classical and basic subject,while defect detection plays an important role in related applications.The traditional object detection algorithm mainly involves two steps.The first is to extract the artificial features designed in the image and the second is to classify the images with the features as input by a classifier.There are two main disadvantages in this method,one is the difficulty of feature engineering,the other is that the performance of the classification model is susceptible to lighting,noise,texture of the part and other conditions.With the vigorous development of deep learn-ing,more and more different detection models have been developed to detect objects,which have achieved ideal results in many fields,especially in competitions.However,those models always demand a mountain of samples and are not good at detecting little objects while most of defections are very small.To solve this problem,this paper takes the normal map as the input of defect de-tection for the first time.We do the defect detecion with a traditional method combined SVM,Adaboost and Cascade and a state-of-the-art performance is acheived.We also take the normal map as input as convolutional neural network,and compare the results with traditional method’s.The main work and innovation of this article could be found as follows:1.We propose a method to acquire the normal map of parts’ surface and design an equipment to acquire the method’s inputs.This equipment includes a camera,LED light bars,a stepper motor and so many other components.All components are constituted into five models so that we could get required information.And we put forward a method to compute the normal maps with the inputs collected by our equipment metioned above.To make up the imperfection of the capture equipment,we add some corrections and enhancements before the normal map’s computation.Those are color correction,white balance correction,distortion correction and light compensation.2.We also do some experiments on traditional object detection methods with the nor-mal maps extracted by us.We extract many diffierent features form the normal maps such as Haar-like,LBP,GLCM,Gradient,HOG and others,and the model’s results and performances with those features as input are evaluated separately.Then we combine those features according to the model’s results and performances.Fur-thermore,a variety of ways are tried to get more data such as rotation and mirroring because of the rareness of defections on the parts.And considering the imbalance of data between the parts with defections and those without defection,we divide the datasets into several subdatasets.For each subdatasets,we use the the method of Adaboost to train a more powerful classifier with the base learner trained by SVM,then we use Cascade to combine those classifier to detect defections.3.At last,we make some experiments with deep learning models.First,we use the sliding windows to divide the original normal map into some smaller windows,and do some data augmentation on those windows as before.We also divide the datasets into some subdatasets and train a model for each subdataset.We combine those moodels with Cascade to detect defections.We also analyse the results and performances of this model comparing to traditional methods’.