Monitoring fishing activities presents several challenges, particularly the need to handle vast amounts of data. Manually reviewing all fishing activities is unfeasible, making the adoption of artificial intelligence (AI) technologies a promising solution. AI technologies can help address these challenges, promote sustainable practices, and improve transparency in fisheries management, thereby ensuring better compliance with regulatory requirements.
In recent years, computer vision algorithms have shown potential for automating fish detection and weight estimation, offering a way to record fishing activities. However, the performance of these algorithms is often hindered by the complexity of the images captured on conveyor belts. Variations in fish appearance, occlusions caused by overlapping fish or debris (e.g., starfish, crabs), and the diverse composition of catches present significant challenges. These factors vary by haul, season, and even within species due to natural variations. Additionally, when these algorithms are applied to new domains, such as different fishing vessels or regions with unfamiliar species, their performance often declines because they were not trained on such diverse datasets. Addressing these issues requires methods to enhance generalisation in out-of-domain scenarios.
Enhancing Generalisation in Computer Vision Algorithms
To cope with data variability, computer vision algorithms must be capable of detecting fish across diverse scenarios. Achieving this goal necessitates extensive training data to account for the wide range of variations. Although cameras on fishing vessels can capture large volumes of images, annotating each one is a time-intensive process requiring expert input for accurate labeling.
To address these challenges, our group focuses on evaluating strategies to improve generalisation with limited training data. The goal is to develop more robust models suitable for large-scale deployment. We plan to explore approaches such as active learning and self-supervised learning to optimise the annotation process and enhance model performance.
Active Learning for Efficient Annotation
Given the vast amount of unlabelled data, active learning can reduce the annotation burden by selecting only the most informative images from the dataset based on uncertainty metrics. These selected images are then annotated and used to (re)train the models. This targeted approach minimises the number of images requiring annotation while ensuring that the selected samples contribute significantly to improving the model’s performance. For example, active learning can prioritise rare or challenging images, such as those containing hard-to-detect fish, to enrich the diversity of the training dataset. This results in a more efficient training process, enabling the model to generalise better with fewer annotated images.
Leveraging Self-Supervised Learning
To complement active learning, self-supervised learning techniques can be employed to utilise unlabelled data effectively. Self-supervised learning generates labels from the data itself, identifying inherent patterns without relying on expert annotations. This method is particularly valuable for pre-training models on extensive unlabelled datasets, such as the images captured by onboard cameras. The pre-trained models can then serve as the foundation for the active learning pipeline, enabling the selection of uncertain images to be more precise and effective. This combined approach can significantly improve the efficiency and accuracy of fish detection models.
Developing more robust models capable of handling diverse variations in data is critical to making automated fish detection systems feasible for large-scale monitoring of fishing activities. By reducing the manual labour required from fishers and providing an efficient means to expand coverage, these advancements can support sustainable fisheries management and improve compliance with control measures.
Article written by Dr. Manuel Córdova, Wageningen University