رضا ترابي مقتدر

Visual inspection of casting parts is considered as a necessary thing in the quality control management system of factories. Today, in order to reduce the effects of human errors and increase the accuracy of detecting apparent defects, intelligent systems based on deep neural networks are being developed rapidly. One of these models is deep learning based on convolutional neural networks. These networks have attracted the attention of researchers due to the automatic extraction of features from the image and the power of appropriate classification of features. In this experimental work, with the aim of creating a simple and cheap method, the use of a deep learning model is presented in order to discover the defective parts in the machining production line. This model is trained by a set of defective and complete images. The part selected in this research is the EF7 engine cylinder block. This part is currently being produced in three production sites of Iran Khodro Company. The selected piece has a complex geometry and the imaging was done without any special hardware facilities and only with the available facilities. In this experience, two general goals have been considered. The first goal in this experiment is to automatically distinguish the images of defective parts from complete parts in such a way that the images of the parts can be classified into two classes of complete and defective parts. The second goal is to automatically detect the location of the defect on the images of the defective parts, so that the location of the defect is displayed on the images with an enclosing rectangle. In order to robustness the deep learning model, different image processing operations have been performed on the images before injecting them into the network. The eva‎luation of the obtained results shows the direct effect of the above operations on the accuracy of model training and the accuracy of the model in detecting defective parts by the trained model. In order to investigate the effect of the type of network selected in the first objective on the speed and accuracy of training, two networks, Alexnet and Resnet 50, have been investigated. The results show that by choosing the retrofitting approach and using Resnet architecture, the classification training accuracy in the first objective is around 98% and the operational accuracy in the industrial environment on new images is around 86%. In order to achieve the second goal, using the YOLO object recognition algorithm, a specialized version of YOLO is presented that has the ability to detect the location of defects on defective images. In this model, the detection accuracy has been improved up to 92% using the retrofitting approach.

يادگيري عميق , بينايي ماشين , بازرسي بصري , عيوب ريخته گري

Deep Learning , Machine Vision , visual inspection , Casting Defects

Reza Torabi Moghtader

Dr. Mohammad Shahbazi