Deep learning model for defect analysis in industry using casting images

Casting is the main backbone of the manufacturing industry in which liquefied metal is put into the desired shape of mold for the reshaping of metal. Hence, casting defect analysis is an essential tool for manufacturers to ensure the quality of their products, reduce costs, and improve customer satisfaction. In this research, an ensemble model is proposed for casting defect analysis by ensembling transfer learning ResNet50 model and proposed Convolution Neural Network (CNN) model. The transfer learning ResNet50 is chosen after analysing the performance of four models i.e. EfficientNetB3, ResNet18, VGG16 and ResNet50. The feature maps extracted from two different models ResNet50 and CNN are weighted ensemble by grid search combination of weights, weight1 and weight2 assigned to these two models respectively to design a hybrid feature map. With the aid of ensemble model, overfitting can be mitigated, which pool the knowledge of numerous models that have been trained on separate portions of the data. The model is implemented using the Kaggle dataset having 7348 images of two different casting classes of defective and non-defective. The proposed ensemble model is simulated and analyzed using three hyper-parameters i.e. optimizers, batch size, and epochs. The proposed ensemble model outperforms the two individual models with the value of optimizer as Adam, batch size as 32, and epochs as 30. The values of precision and accuracy for the ensemble model come out to be 99.89% and 98.18% respectively.