Search

Abstract Identifying and counting fish individuals on photos and videos is a crucial task to cost-effectively monitor marine biodiversity, yet it remains difficult and time-consuming. In this paper, we present a method to assist the identification of fish species on underwater images, and we compare our model performances to human ability in terms of speed and accuracy. We first tested the performance of a convolutional neural network (CNN) trained with different photographic databases while accounting for different post-processing decision rules to identify 20 fish species. Finally, we compared the performance of species identification of our best CNN model with that of humans on a test database of 1197 fish images representing nine species. The best CNN was the one trained with 900,000 images including (i) whole fish bodies, (ii) partial fish bodies and (iii) the environment (e.g. reef bottom or water). The rate of correct identification was 94.9%, greater than the rate of correct identification by humans (89.3%). The CNN was also able to identify fish individuals partially hidden behind corals or behind other fish and was more effective than humans to identify fish on smallest or blurry images while humans were better to identify fish individuals in unusual positions (e.g. twisted body). On average, each identification by our best CNN using a common hardware took 0.06 s. Deep Learning methods can thus perform efficient fish identification on underwater images and offer promises to build-up new video-based protocols for monitoring fish biodiversity cheaply and effectively. Highlights • Comparison between human experts and Deep Learning based method • Assessing the importance of post treatment with DL based methods • Importance of database adjustements • Real-life use cases for applied DL methods for fish identification [ABSTRACT FROM AUTHOR]

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. • First comparison between Deep and Few-Shot Learning for fish identification • Few-Shot outperforms Deep Learning in situations of small image dataset. • Classification accuracy is asymptotic function of training dataset size in our case study. • Asymptotic accuracy of 76% with 1153 images per class in Deep Learning • Asymptotic accuracy of 60% with 7 images per class in Few-shot Learning [ABSTRACT FROM AUTHOR]