Underwater cameras are widely used to monitor marine biodiversity, and the trend is increasing due to the availability of cheap action cameras. The main bottleneck of video methods now resides in the manual processing of images, a time-consuming task requiring trained experts. Recently, several solutions based on Deep Learning (DL) have been proposed to automatically process underwater videos. The main limitation of such algorithms is that they require thousands of annotated images in order to learn to discriminate classes (here species). This limitation implies two issues: 1) the annotation of hundreds of common species requires a lot of efforts 2) many species are too rare to gather enough data to train a classic DL algorithm. Here, we propose to explore how few-shot learning (FSL), an emerging research field, could overcome DL limitations. Few-shot learning is based on the principle of training a Deep Learning algorithm on “how to learn a new classification problem with only few images”. In our case-study, we assess the robustness of FSL to discriminate 20 coral reef fish species with a range of training databases from 1 image per class to 30 images per class, and compare FSL to a classic DL approach with thousands of images per class. We found that FSL outperform classic DL approach in situations where annotated images are limited, yet still providing good classification accuracy.

水下摄像机被广泛用于监测海洋生物多样性，并且由于廉价运动摄像机的出现，这种趋势正在增加。现在，视频方法的主要瓶颈在于图像的手动处理，这是一项耗时的任务，需要训练有素的专家。最近，已经提出了几种基于深度学习（DL）的解决方案来自动处理水下视频。这种算法的主要局限性在于，它们需要成千上万个带注释的图像，才能学会区分类别（此处为物种）。这种局限性意味着两个问题：1）对数百种常见物种进行注释需要付出很多努力2）许多物种太稀少，以至于无法收集足够的数据来训练经典的DL算法。在这里，我们建议探索少拍学习（FSL）这一新兴的研究领域如何克服DL的局限性。少量学习基于训练深度学习算法的原理，即“如何只用很少的图像学习新的分类问题”。在我们的案例研究中，我们通过一系列训练数据库（从每班1张图像到每班30张图像）的训练数据库，评估了FSL区分20种珊瑚礁鱼类的鲁棒性，并将FSL与经典DL方法进行比较，每级上千张图像。我们发现，在带注释的图像有限的情况下，FSL优于传统的DL方法，但仍提供了良好的分类精度。我们使用一系列训练数据库（从每班1张图像到每班30张图像）评估FSL区分20种珊瑚礁鱼类的鲁棒性，并将FSL与经典DL方法（每班数千张图像）进行比较。我们发现，在带注释的图像有限的情况下，FSL优于传统的DL方法，但仍提供了良好的分类精度。我们使用一系列训练数据库（从每班1张图像到每班30张图像）评估FSL区分20种珊瑚礁鱼类的鲁棒性，并将FSL与经典DL方法（每班数千张图像）进行比较。我们发现，在带注释的图像有限的情况下，FSL优于传统的DL方法，但仍提供了良好的分类精度。