An effective microscopic image augmentation approach

There are too many types of Chinese medicinal herbs (CMH) and it is difficult to collect microscopic images, which naturally leads to the problem of small sample size. In addition, CMH also has some scarcity characteristic, with the proportion of certain cells as low as 0.5%. This leads to the failure of deep learning models, and even few-shot learning methods are difficult to solve effectively. Expanding the data scale of rare features is one of the effective strategies. To address this challenge, we propose an effective microscopic image augmentation approach for few-shot learning (MIAA-FSL). The approach consists of two aspects: first, we design the conditionally guided microscopic image generation model (CGMIGM), which combines the denoising diffusion probabilistic models (DDPM) based conditional guidance technique to efficiently generate rare features and thus alleviate the class imbalance problem. Second, we introduce the semi-supervised learning data augmentation model (SSLDAM), which integrates semi-supervised image processing and pseudo-label generation techniques to effectively overcome the issues of damage, blurriness, and difficulty in discernment in microscopic images, making otherwise unusable images usable. The experimental results show that the MIAA-FSL improves the identification accuracy by 24% on average compared with the Microscope Image Recognition + DDPM (MIR+DDPM) approach, especially in the identification of rare features, the accuracy is significantly improved from 45.5% to 87.0%, which effectively mitigates the problem of object detection with few samples.