With the applications of deep learning in computer vision and natural language processing continuing to deepen, deep neural networks have obtained better processing results than traditional algorithms for complex situations such as bubbles, noise, and uneven illumination. However, deep neural networks usually require a large number of training data, which is difficult to collect in traditional fields such as biology, materials, physics, and medicine, where the data collection is time-consuming and expensive. Therefore, the importance of few-shot learning is increasing daily. Few-shot learning is a strategy to complete deep neural model training based on a small dataset, and the specific implementation includes data augmentation, migration learning, meta-learning, etc. In this paper, the...[ Read more ]

With the applications of deep learning in computer vision and natural language processing continuing to deepen, deep neural networks have obtained better processing results than traditional algorithms for complex situations such as bubbles, noise, and uneven illumination. However, deep neural networks usually require a large number of training data, which is difficult to collect in traditional fields such as biology, materials, physics, and medicine, where the data collection is time-consuming and expensive. Therefore, the importance of few-shot learning is increasing daily. Few-shot learning is a strategy to complete deep neural model training based on a small dataset, and the specific implementation includes data augmentation, migration learning, meta-learning, etc. In this paper, the problem of insufficient data in the target domain is solved by data augmentation methods for digital PCR, real-time PCR, and bacterial colony counting. Traditional algorithms have limited accuracy for these fields because of the complicated interference, and effective data augmentation methods can greatly reduce the data collection cost of deep learning. Compared with the traditional algorithms, the dPCR, real-time PCR, and bacterial colony images, the deep learning methods based on the few-shot learning strategy proposed in this paper improve the accuracy from 64%, 49.1%, and 4.4% to 99%, 93.9%, and 97.4%, respectively. The proposed data augmentation methods, Random Background Cover Method (RBTM) and Random Cover Targets Algorithm (RCTA) can effectively augment data by more than 200 times while reducing data annotation time by more than 70%. The application of the few-shot learning strategy can bring more possibilities for the implementation of deep learning in traditional fields, which provides a new approach to improve the accuracy of complex interference images and hugely reduce the data collection costs.