Microscopic algae images originate from various sources such as microscopes, specialized imaging tools, online platforms, and different formats. For scientists tracking harmful algal blooms, this variety presents a challenge. An AI trained on one image type might not perform well on others. To address this, the patent proposes a multi-source algae detection system tailored for this mismatch. It begins with an automated tool that collects images of targeted algae, filtering out unsuitable ones based on cell content, clarity, and interference. The selected images are labeled to create a source-domain dataset. An unlabeled target-domain dataset is then added, and transfer learning with Faster R-CNN is applied. The system adapts network parameters and employs multi-kernel maximum mean discrepancy to minimize differences between source and target domains. The aim is a versatile model capable of recognizing and classifying various algae across diverse sources and formats. This is vital because harmful blooms can harm freshwater ecosystems and involve toxin-producing species. Identifying these organisms is essential for monitoring. The broader AI challenge is that valuable data exists but isn’t collected under a single protocol. Instead of dismissing incompatible images, this method seeks to bridge the gap, enabling AI to learn from fragmented information.
Recognizing algae across images that were never made to match
Patent number: CN 115311657 B
Inventor(s): J. Li, A. Yuan, B. Wang, H. Zou, and J. Wang
Citation: J. Li, A. Yuan, B. Wang, H. Zou, and J. Wang, “多源藻类图像目标检测方法、系统、电子设备及存储介质 [Multi-source algae-image object-detection method, system, electronic device and storage medium],” China Patent CN 115311657 B, Jan. 5, 2024.
