Research On General Detection Method Of Multi-type Defects Of Multi-type Iron Cans

Defect detection is an important process in the production process of the product.Its results can reflect the existing problems in the production process,which plays an important role in improving product quality and maintaining the reputation of enterprises.Compared with manual product defect detection,the automatic product defect detection method based on computer vision has the advantages of high efficiency and low cost.At the same time,the machine can replace the workers in high temperature,poisonous gas,radiation and other harsh environments to carry out defect detection work.At present,there are two product defect detection methods based on computer vision.One is a detection method based on digital image processing.This method uses image alignment and image contrast technology to detect the defect location of the product,and then distinguish the defect type by the image characteristics of the defect area(such as gray pixel area).It has the characteristics of fast detection speed and high accuracy,but its commonality and expandability are poor.For multi-type defects of products,professionals must re-select the features used to distinguish the defect category.It is more suitable for use in defect detection of single product.The other method is a product defect detection method based on deep learning.This method trains a neural network model that detects multiple types of defects in multiple products.It has the characteristics of end-to-end detection and good universality,and is suitable for the detection of multiple types of defects of multiple products.Aiming at the extremely high universality and detection accuracy requirements of defect detection task of multi-type defects of multi-type iron cans of enterprise,this paper carrys out researches on the general detection method of multi-type defects of multi-type iron cans based on deep learning.The main results are as follows:(1)This paper analyzes the reasons why the current state-of-the-art deep learning-based object detection algorithms perform poorly in the defect detection task of multi-type defects of multi-type iron cans,and designs a new product defect detection algorithm,called FPP(Feature Pair Pyramid),based on the analysis results.In FPP algorithm,a new feature extraction method and three types of defect detection optimization methods are proposed.The experimental results show that the FPP defect detection algorithm performs better than the current advanced deep learning object detection algorithm in the detection of multi-type defects of multi-type iron cans.The three optimization methods proposed also greatly improve the accuracy of defect detection,especially the accuracy of small-size defect detection.(2)Aiming at the characteristics of high cost and time-consuming labeling of data sets in deep learning,an automatic labeling method for product defect detection data sets in deep learning is proposed.This method integrates image alignment,image contrast and prior prediction methods for defect location and defect category labeling.Under certain conditions,the labeling accuracy of this data set automatic labeling method can reach 90.44%.(3)Aiming at the problem of time-consuming retraining of defect detection models,a method for accelerating the training of defect detection models is proposed.This method uses hard sample mining methods,which can retrain new models in a short time using a small number of new product samples,thus saving the time cost.