Siamese Basis Function Networks for Data-Efficient Defect Classification in Technical Domains

Training deep learning models in technical domains is often accompanied by the challenge that although the task is clear, insufficient data for training is available. Additional to that, often also no similar source-datasets are available which can be used for transfer-learning to reduce the need for data in the target-domain. In this work, a novel approach based on the combination of siamese networks and radial basis function networks is proposed where the siamese networks serve as effective feature extractors. The architecture performs data-efficient classification without pretraining by measuring the distance between images in the semantic space in a data- efficient manner. The so called SBF-Net structure is developed and tested on three technical as well as two non-technical datasets. The architecture shows superior performance for all data sets, especially when only small data is available for training. The approach significantly outperforms existing ResNet50 and ResNet100 architectures when only 3, 5, 10 and 20 data points per class are available. Also, in data-setups where 75% of the data is used for training, the model yields the same performance as state-of-the-art-models. The main contribution of this work is a model that works particularly data efficient with small amounts of data without making prior constraints.