Explainable patch-wise crack detection and image classification on the egg shell dataset

Abstract

From the egg manufacturing to distributing, the most important part is to sort the sellable eggs and the non-sellable eggs. Traditionally classifying process has been done one by one by the egg experts. However, nowadays, all the processes from manufacturing to distributing has been automatized and following such trend, the need to automatize the egg classification process is emerging. To effectively classify the eggs into categories such as cracked, blood spotted, bleached or deformity, the main task is to analyze the characteristics of each category on the images of egg shells taken by the cameras. Existing researches on automatizing egg classification process using egg images can be divided into two groups: crack detection and classification research based on the traditional image processing methods and egg classification research based on the deep learning. The researches based on the traditional image processing methods used the methods such as threshold binarization or edge detection to detect the cracks on the egg shell to classify the eggs. However, the image processing methods inherently are sensitive on the input parameters that even in the same dataset, noises come up in the output images or some parts of the cracks are lost. Also, this characteristic makes it hard to implement the same image processing process on the other dataset. The researches based on the deep learning need to provide explanations on their models to convince others about their models’ work, but the existing deep learning-based egg classification researches often failed on providing enough explainability. It is a problem to be solved since to have analysis on the wrong bias of the model and to compensate it, human interpretable explanation on the model is required. In this research, we propose a patch-wise crack detection and image classification method on the egg shell dataset to compensate the limitations of existing works. The proposed method is trained on the sliced image and put the inference results from each slice together to detect the cracks more elaborately. Also, this method can classify the eggs into normal, crack, and latent crack which is a new category we found during the research in a prioritized manner, and can provide enough explanation on the model.