Limitations
Although SealNet produced promising results, there are still limitations that need to be addressed. First, our SealNet software still requires some manual work during the data collection process—after running the automatic face detector, researchers are still required to manually locate the eyes, nose and mouth in order for the program to automatically align and chip the seal faces. Thus, one possible improvement that we can implement in the future is to add a landmark detector to be used in conjunction with the face detector. Secondly, to generate training data, researchers must manually group multiple face chips belonging to the same individual. Not only is this process laborious, it may be also error-prone. A more sustainable approach would be to implement a classifier; however, researchers would still be required to manually check if the classification is accurate.
Although SealNet does well in closed-set classification, open-set verification performance could be dramatically improved by reducing the similarity scores between such seals. This success could be achieved in two ways. First, changes in preprocessing methodology could yield greater differentiation between very similar seals by removing lighting and other environmental effects without affecting color pattern information the way grayscaling does. Second, changes to our model architecture could improve such performance. However, the inherent complexity in any attempt to leverage specificity, while simultaneously avoiding overfitting, presents a difficult balance which all recognition models struggle to strike. Thus, the best approach to this problem would be to maximize the quantity and quality of information available to the model through preprocessing improvements prior to making changes to the CNN architecture itself.