Search

Data augmentation is an established technique in computer vision to foster the generalization of training and to deal with low data volume. Most data augmentation and computer vision research are focused on everyday images such as traffic data. The application of computer vision techniques in domains like marine sciences has shown to be not that straightforward in the past due to special characteristics, such as very low data volume and class imbalance, because of costly manual annotation by human domain experts, and general low species abundances. However, the data volume acquired today with moving platforms to collect large image collections from remote marine habitats, like the deep benthos, for marine biodiversity assessment and monitoring makes the use of computer vision automatic detection and classification inevitable. In this work, we investigate the effect of data augmentation in the context of taxonomic classification in underwater, i.e., benthic images. First, we show that established data augmentation methods (i.e., geometric and photometric transformations) perform differently in marine image collections compared to established image collections like the Cityscapes dataset, showing everyday traffic images. Some of the methods even decrease the learning performance when applied to marine image collections. Second, we propose new data augmentation combination policies motivated by our observations and compare their effect to those proposed by the AutoAugment algorithm and can show that the proposed augmentation policy outperforms the AutoAugment results for marine image collections. We conclude that in the case of small marine image datasets, background knowledge, and heuristics should sometimes be applied to design an effective data augmentation method.

Underwater computer vision plays a vital role in ocean research, enabling autonomous navigation, infrastructure inspections, and marine life monitoring. However, the underwater environment presents unique challenges, including color distortion, limited visibility, and dynamic light conditions, which hinder the performance of traditional image processing methods. Recent advancements in deep learning (DL) have demonstrated remarkable success in overcoming these challenges by enabling robust feature extraction, image enhancement, and object recognition. This review provides a comprehensive analysis of cutting-edge deep learning architectures designed for underwater object detection, segmentation, and tracking. State-of-the-art (SOTA) models, including AGW-YOLOv8, Feature-Adaptive FPN, and Dual-SAM, have shown substantial improvements in addressing occlusions, camouflaging, and small underwater object detection. For tracking tasks, transformer-based models like SiamFCA and FishTrack leverage hierarchical attention mechanisms and convolutional neural networks (CNNs) to achieve high accuracy and robustness in dynamic underwater environments. Beyond optical imaging, this review explores alternative modalities such as sonar, hyperspectral imaging, and event-based vision, which provide complementary data to enhance underwater vision systems. These approaches improve performance under challenging conditions, enabling richer and more informative scene interpretation. Promising future directions are also discussed, emphasizing the need for domain adaptation techniques to improve generalizability, lightweight architectures for real-time performance, and multi-modal data fusion to enhance interpretability and robustness. By critically evaluating current methodologies and highlighting gaps, this review provides insights for advancing underwater computer vision systems to support ocean exploration, ecological conservation, and disaster management.

SCoralNet (based on Faster R-CNN) is a new underwater coral detection framework that has been proposed to automatically localize and identify distinct coral species in images, allowing for fast and detailed annotation. Monitoring the coverage and abundance of underwater corals typically involves the annotation and processing of large amounts of underwater coral images. However, manually annotating a large number of images is time-consuming and labor-intensive, and CNN classifiers only provide simple classification annotations without capturing the images' finer details. SCoralNet's detection performance is improved by incorporating dilated convolutions into the backbone network. To successfully capture multi-scale and multi-level information from coral targets, a neck network called NASFPN is placed between the backbone and the detecting head. Seesaw Loss is used to reduce the impact of the dataset's long-tailed distribution on SCoralNet's classifier accuracy. CIoU loss is used to optimize the bounding box regression method. During inference, Soft-NMS is applied to suppress redundant coral detection boxes. To assess SCoralNet's effectiveness, a dataset called Coral-soft was developed using real-world photos of common soft coral species from the Sanya region of China. SCoralNet outperformed the original Faster R-CNN model on the Coral-soft dataset, with a 45.68% gain in mean average precision (mAP) and a 59.2% increase in mAP75. Furthermore, SCoralNet outperformed some advanced models in terms of overall performance. [ABSTRACT FROM AUTHOR]

Fish assessment and monitoring are important for the development of a modern aquatic ecosystem. Fish are a vital part of the marine and freshwater environments. Morphological and computational details of fish, such as size, shape, and position, are important in fish observation and fisheries. Typically, manual, or low-efficient techniques are used to acquire fish details. However, existing typical methods are usually time-consuming, less accurate, and resource-intensive. Computer-aided methods are crucial for intelligent and automatic fish assessment. Two novel networks, namely parallel feature fusion-based segmentation network (PFFS-Net) and progressive information fusion-based segmentation network (PIFS-Net), were developed for pixel-wise fish segmentation. PFFS-Net is a base network that uses parallel feature fusion to achieve a better segmentation performance. PIFS-Net is the final model of this work and uses a progressive spatial feature fusion (SFF) mechanism to enhance segmentation accuracy. PIFS-Net also employs rapid feature reduction and pre-prediction low-level information fusion blocks to further boost performance.The proposed models were evaluated using the following three publicly available databases: semantic segmentation of underwater imagery (SUIM), DeepFish, and Large-scale fish. The proposed networks outperformed the state-of-the-art methods in challenging underwater conditions with superior computational efficiency. PIFS-Net needs only 2.02 million trainable parameters for its complete training. Automatic and accurate fish segmentation can be a major step towards an intelligent aquatic ecosystem. The codes of our algorithms and trained models are available on Github.

Fish assessment and monitoring are important for the development of a modern aquatic ecosystem. Fish are a vital part of the marine and freshwater environments. Morphological and computational details of fish, such as size, shape, and position, are important in fish observation and fisheries. Typically, manual, or low-efficient techniques are used to acquire fish details. However, existing typical methods are usually time-consuming, less accurate, and resource-intensive. Computer-aided methods are crucial for intelligent and automatic fish assessment. Two novel networks, namely parallel feature fusion-based segmentation network (PFFS-Net) and progressive information fusion-based segmentation network (PIFS-Net), were developed for pixel-wise fish segmentation. PFFS-Net is a base network that uses parallel feature fusion to achieve a better segmentation performance. PIFS-Net is the final model of this work and uses a progressive spatial feature fusion (SFF) mechanism to enhance segmentation accuracy. PIFS-Net also employs rapid feature reduction and pre-prediction low-level information fusion blocks to further boost performance. The proposed models were evaluated using the following three publicly available databases: semantic segmentation of underwater imagery (SUIM), DeepFish, and Large-scale fish. The proposed networks outperformed the state-of-the-art methods in challenging underwater conditions with superior computational efficiency. PIFS-Net needs only 2.02 million trainable parameters for its complete training. Automatic and accurate fish segmentation can be a major step towards an intelligent aquatic ecosystem. The codes of our algorithms and trained models are available on Github. [ABSTRACT FROM AUTHOR]