A Comparative Study on the Use of Raw and Filtered Images for Multi-class Image Classification

Abstract

Manufacturing processes have since then developed to accommodate the ever-growing consumer demand for a wide range of products. Being so, for businesses to adapt, they must engage in the manufacturing and selling of various types of products. Often times in the factory, one would oversee the overall manufacturing processes of several products. This has been proven to be workable, yet, at times, inefficient. Multi-class image classification may serve to be the answer for increasing efficiency in monitoring the overall manufacturing process. To demonstrate the technology and apply it in an electronic device manufacturing setting, the researchers created their own dataset that consists of 4 different classes (with each class having 1000 images -- 1000 for training, 518 for validation, and 5 for testing). Following the Convolutional Neural Network (CNN) model, the researchers made use of the Python programming language to execute the multi-class image classification program. It was investigated if there would be any difference in the model and output accuracy if the images that would be fed to the program will undergo no filtering and gaussian filtering. Training and testing were performed for a total of 5 times for each set, and training loss, training accuracy, validation loss, validation accuracy, and actual accuracy were recorded. The mean values of the aforementioned show that Gaussian filtered images yielded much more favorable results as actual accuracy was at 95% -- which is 5% higher than the actual accuracy for the set of raw images.