CleanCam: a labelled image dataset for camera-cleaning decisions in aquaculture monitoring
Minh Khoa Nguyen, Tuan Anh Hoang, Tran, Nam Nhat Anh, Tran, Nam Nguyet Anh, Minh Hoang Pham, Van Khoi Phan, Nguyen, Van Dinh, Dinh, Van Dung, Do, Danh Cuong
CleanCam is a benchmark dataset for underwater camera-viewport fouling severity assessment in aquaculture. It distinguishes material attached to the camera viewport from water-column degradation, including turbidity, haze, suspended particles, lighting variation, and low contrast. These conditions can appear similar in individual frames but imply different operational responses.CleanCam v2.0.0 contains 22,497 RGB JPEG images at 3072 × 2048 pixels. The dataset includes 18,897 real images sampled every 10 seconds from approximately 53 hours of valid underwater video collected over 20 effective collection days using two fixed cameras in golden pompano (Trachinotus blochii) rearing tanks at the Research Institute for Aquaculture No.1 in Hai Phong, Viet Nam. It also includes 3,600 split-consistent synthetic images for Levels 3–5 to support controlled augmentation and stress-testing experiments.Images are annotated using a five-level ordinal severity protocol focused specifically on viewport fouling. Temporal context was used to distinguish persistent surface deposits from transient water-column effects. Clean-but-turbid frames remain at Level 1, while higher levels require stable evidence of material attached to the viewport.This release includes: real and synthetic image folders organized by severity level; master, real-image, and synthetic-image metadata tables; deterministic capture-disjoint train, validation, and test splits; image-level synthetic provenance, including parent images, transformation parameters, opacity, blockage estimates, assets, and random seeds; 11 documented RGBA deposit assets assigned to disjoint train, validation, and test pools; a complete data dictionary, label taxonomy, acquisition-condition documentation, deposit-asset documentation, and quickstart guide; release summaries and a SHA-256 file manifest; and release-building and validation scripts. The metadata describe image identifiers, file paths, labels, camera nodes, recording sessions, dates, elapsed seconds, capture identifiers, dimensions, and checksums. Synthetic records additionally preserve parent-image linkage and generator parameters.CleanCam can support ordinal classification, camera-cleaning decision studies, leakage-aware benchmarking, synthetic augmentation, camera-health monitoring, underwater robotics, and image-stream quality control in long-term aquatic monitoring systems.Source code and reproducibility materials:https://github.com/khoa288/CleanCam/releases/tag/v2.0.0