Sea cage inspection is a crucial task in marine aquaculture for both controlling the quality in the fish development process and the environmental impact of these facilities. Net damage in sea cages can cause fish to escape from fish farms to the open sea, producing serious economic losses and compromising the surrounding marine ecosystem. Currently, the most sophisticated inspection processes used in the industry rely on human visual inspection of underwater videos captured with tele-operated underwater ROV (Remotely Operated Vehicles). This process is tedious, time consuming, imprecise, highly dependant on the operator level of expertise, and has low verifiability.

This article presents a comprehensive algorithm for automatic net damage detection in sea cages for aquaculture oriented to on-ROV processing and real-time processing. The proposed approach takes a video stream from an on-ROV camera, segments the image frames to separate the net of the sea cage from the background, and applies noise reduction tuned for underwater conditions. Then, to perform net damage detection, the mesh net hole areas are analyzed for detecting outliers that represent potential damage. Finally, holes neighboring to the outliers are analyzed to reduce perspective errors, and a spatial-temporal criterion using tracking is applied, to reduce the chance of false positives. The algorithm was first tested using a set of public images for comparison with state-of-the-art approaches and repeatability of the tests, and then using a dataset of real ROV inspection sequences to evaluate its effectiveness in real-world scenarios. Results show that our approach presents high levels of accuracy even for adverse scenarios and is adequate for real-time processing in embedded platforms.

海水网箱检查是海水养殖中的一项重要任务，既可以控制鱼类发育过程中的质量，也可以控制这些设施对环境的影响。海网箱中的网箱损坏会导致鱼类从养鱼场逃到公海，造成严重的经济损失并危及周围的海洋生态系统。目前，行业中使用的最复杂的检查流程依赖于人工目视检查使用遥控水下 ROV（遥控车辆）拍摄的水下视频。这个过程繁琐、耗时、不精确，高度依赖于操作员的专业水平，并且可验证性低。

本文提出了一种面向 ROV 处理和实时处理的水产养殖海网箱自动网损自动检测的综合算法。所提出的方法从 ROV 摄像机获取视频流，分割图像帧以将海笼网与背景分开，并应用针对水下条件调整的降噪。然后，为了执行网络损坏检测，分析网格网孔区域以检测代表潜在损坏的异常值。最后，分析与异常值相邻的孔以减少透视误差，并应用使用跟踪的时空标准来减少误报的机会。该算法首先使用一组公共图像进行测试，以便与最先进的方法和测试的可重复性进行比较，然后使用真实 ROV 检查序列的数据集来评估其在现实场景中的有效性。结果表明，即使在不利情况下，我们的方法也能提供高水平的准确性，并且足以在嵌入式平台中进行实时处理。