Visual Inspection of Casting Product Using Convolutional Neural Network (CNN)

Abstract

At the end of manufacturing of casting process, it is crucial to have a process which inspect if it has defects or not. Conventional method for inspection uses naked eyes which is prone to misclassify the product. The objectives for this study are to identify defects by using Convolutional Neural Network (CNN) simultaneously reduce the number of misclassification and to evaluate the performance of the CNN by comparing it with other image processing method. This approach is only applicable if the product is circle in shape and by using Python programming language for the codes. CNN is machine learning that specific in image processing. To build a fully functioning algorithm, the researcher has used google colab as the notebook and it should consist of 3 layers called convolutional layer where it extract all the feature contains in the raw image, pooling layer will reduce the size of the feature map to lessen the complexity for the next layers, and fully-connected layer is to combine all features to become an output. There are six criteria used when comparing CNN and ResNet. Final results prove that ResNet has a better performance as it is much simple to construct the algorithm because there is only 5 line of codes to import all libraries into the notebook while in CNN, the researcher need to start from scratch. The accuracy for determining the product are not much different compared to CNN. The accuracy for ResNet and CNN are 99.86% and 99.72%. For future researcher it is recommended to do a research related to Recurrent neural Network (RNN). RNN is commonly used in speech recognition.