Bacterial colony counting is a time consuming but important task for many fields, such as food quality testing and pathogen detection, which own the high demand for accurate on-site testing. However, bacterial colonies are often overlapped, adherent with each other, and difficult to precisely process by traditional algorithms. The development of deep learning has brought new possibilities for bacterial colony counting, but deep learning networks usually require a large amount of training data and highly configured test equipment. The culture and annotation time of bacteria are costly, and professional deep learning workstations are too expensive and large to meet portable requirements. To solve these problems, we propose a lightweight improved YOLOv3 network based on the few-shot learning strategy, which is able to accomplish high detection accuracy with only five raw images and be deployed on a low-cost edge device. Compared with the traditional methods, our method improved the average accuracy from 64.3% to 97.4% and decreased the False Negative Rate from 32.1% to 1.5%. Our method could greatly improve the detection accuracy, realize the portability for on-site testing, and significantly save the cost of data collection and annotation over 80%, which brings more potential for bacterial colony counting.

With the global pandemic of infectious diseases, the demand for accurate nucleic acid detection is daily increasing. The traditional threshold-based algorithms are adopted as the mainstream for processing the images of digital polymerase chain reaction (dPCR) now, but they are facing huge challenges on complex problems such as irregular noise, uneven illumination, and the lack of data. So, this paper proposed a novel few-shot learning method based on our improved YOLOv3 model with fast processing speed and high accuracy to deal with complicated situations. Besides, to reduce the requirement of the large training dataset and annotation time of deep neural networks, we proposed the Random Background Transfer Method (RBTM) and Source Traceability Annotation Method (STAM) as the data augmentation and annotation method separately, which exploit the prior knowledge of the data and successfully realized the few-shot learning. Bases on the domain knowledge of dPCR images, our method could effectively augment images and reduce the labeling time by 70% while retaining the visually prominent features and improves the detection accuracy from 63.96% of the traditional threshold-based algorithm to as high as 98.98%. With the optimal processing speed and accuracy, our method is the state-of-art strategy for the detection of dPCR images now.

We designed a silicon-based fast-generated static droplets array (SDA) chip and developed a rapid digital polymerase chain reaction (dPCR) detection platform that is easy to load samples for fluorescence monitoring. By using the direct scraping method for sample loading, a droplet array of 2704 microwells with each volume of about 0.785 nL can be easily realized. It was determined that the sample loading time was less than 10 s with very simple and efficient characteristics. In this platform, a pressurized thermal cycling device was first used to solve the evaporation problem usually encountered for dPCR experiments, which is critical to ensuring the successful amplification of templates at the nanoliter scale. We used a gradient dilution of the hepatitis B virus (HBV) plasmid as the target DNA for a dPCR reaction to test the feasibility of the dPCR chip. Our experimental results demonstrated that the dPCR chip could be used to quantitatively detect DNA molecules. Furthermore, the platform can measure the fluorescence intensity in real-time. To test the accuracy of the digital PCR system, we chose three-channel silicon-based chips to operate real-time fluorescent PCR experiments on this platform.

As an important part of the industrialization process, fully automated instrument monitoring and identification are experiencing an increasingly wide range of applications in industrial production, autonomous driving, and medical experimentation. However, digital instruments usually have multi-digit features, meaning that the numeric information on the screen is usually a multi-digit number greater than 10. Therefore, the accuracy of recognition with traditional algorithms such as threshold segmentation and template matching is low, and thus instrument monitoring still relies heavily on human labor at present. However, manual monitoring is costly and not suitable for risky experimental environments such as those involving radiation and contamination. The development of deep neural networks has opened up new possibilities for fully automated instrument monitoring; however, neural networks generally require large training datasets, costly data collection, and annotation. To solve the above problems, this paper proposes a new instrument monitoring method based on few-shot learning (FLIMM). FLIMM improves the average accuracy (ACC) of the model to 99% with only 16 original images via effective data augmentation method. Meanwhile, due to the controllability of simulated image generation, FLIMM can automatically generate annotation information for simulated numbers, which greatly reduces the cost of data collection and annotation.