One-Shot Learning for Custom Identification Tasks; A Review
Niall O’Mahony, Sean Michael Campbell, Anderson Luiz De Carvalho, Lenka Krpálková, Gustavo Velasco-Hernandez, Suman Harapanahalli, Daniel Riordan, Joseph L. Walsh · Procedia Manufacturing · 2019
Deep Learning has great achievements in computer vision for various classification and regression tasks. The automation of tasks such as component sorting, bin-picking and anomaly detection may be of great use in the process industry. However, most machine learning-based object categorization algorithms require training on hundreds or thousands of images and very large datasets. The requirement for large training datasets presents a barrier to the adoption of deep learning methodologies in many custom object classification tasks. For example, in defect detection, positive instances of a defect, take for instance a tank leakage, may seldom occur and therefore creating a dataset of sufficient size for conventional deep learning procedures is not always possible. One-shot learning aims to learn information about object categories from only a handful of labelled examples per category. One-shot learning has received the most attention in face-recognition and person re-identification (re-id) tasks due to their potential practical applications in surveillance security. This research will review these one-shot learning methodologies and investigate how they may be transferred to other domains. Concepts such as Siamese Networks and triplet loss which are commonly used for one-shot learning will be examined. Challenges such as variations in illumination conditions, object pose, camera resolution and partial occlusion will be discussed. Finally, the implications and advantages of deploying such techniques to practical applications in the process industry will be analysed.