Self-supervised Learning Approach to Detect Corrosion Products in Biofilm images

Detection of microbially influenced corrosion (MIC) products in biofilm images is an important problem in material science and engineering. Neural network models that can accurately detect biofilm images having corrosion products are likely to have significant and broad implications in the study and development of materials. However, generating and annotating biofilm image datasets in sufficient volumes is a major impediment in developing such networks. A self-supervised learning approach is presented for automatically detecting MIC products on metal surfaces based on analyses of Scanning Electron Microscope (SEM) biofilm images. The proposed approach uses a Simple Siamese (SimSiam) architecture to learn visual image representations from an unlabeled set of biofilm images, which is then fine-tuned using a scarce set of labeled images to build a model to detect biofilms containing corrosion products. The architecture generates two different augmented versions of the input images and learns the representations using an encoder that uses ResNet backbone. The architecture aims to minimize the negative cosine similarity of the outputs from the encoders and hence learns the representations of the images as both augmented versions belong to the same image. In order to improve the dataset quality and volume, input images are contrast enhanced, scaled, and overlapping image patches are generated and used to learn representations and fine-tuning. The performance of the models are analyzed using precision and recall metrics for patches of varying sizes. An overall accuracy of 62% was obtained for the classification of corrosion in the images after finetuning the model with scarce labeled dataset. Our results show that the models built using the proposed self-supervised learning approach can successfully detect corrosion products in biofilm images and that the performance of the models successively improves with increase in the patch size.