It is important for marine scientists and conservationists to frequently estimate the relative abundance of fish species in their habitats and monitor changes in their populations. As opposed to laborious manual sampling, various automatic computer-based fish sampling solutions in underwater videos have been presented. However, an optimal solution for automatic fish detection and species classification does not exist. This is mainly because of the challenges present in underwater videos due to environmental variations in luminosity, fish camouflage, dynamic backgrounds, water murkiness, low resolution, shape deformations of swimming fish, and subtle variations between some fish species. To overcome these challenges, we propose a hybrid solution to combine optical flow and Gaussian mixture models with YOLO deep neural network, an unified approach to detect and classify fish in unconstrained underwater videos. YOLO based object detection system are originally employed to capture only the static and clearly visible fish instances. We eliminate this limitation of YOLO to enable it to detect freely moving fish, camouflaged in the background, using temporal information acquired via Gaussian mixture models and optical flow. We evaluated the proposed system on two underwater video datasets i.e., the LifeCLEF 2015 benchmark from the Fish4Knowledge repository and a dataset collected by The University of Western Australia (UWA). We achieve fish detection F-scores of 95.47% and 91.2%, while fish species classification accuracies of 91.64% and 79.8% on both datasets respectively. To our knowledge, these are the best reported results on these datasets, which show the effectiveness of our proposed approach.

对于海洋科学家和保护主义者而言，重要的是经常估计其栖息地中鱼类的相对丰富度并监测其种群的变化。与费力的手动采样相反，在水下视频中提出了各种基于计算机的自动鱼采样解决方案。但是，不存在用于自动鱼类检测和物种分类的最佳解决方案。这主要是由于水下视频中存在的挑战，这些挑战是由于亮度，鱼伪装，动态背景，水沉闷，分辨率低，游泳鱼的形状变形以及某些鱼种之间的细微变化等环境变化引起的。为克服这些挑战，我们提出了一种混合解决方案，将光流和高斯混合模型与YOLO深层神经网络相结合，在不受限制的水下视频中对鱼类进行检测和分类的统一方法。基于YOLO的对象检测系统最初用于仅捕获静态且清晰可见的鱼实例。我们消除了YOLO的这一限制，使其能够使用通过高斯混合模型和光流获得的时间信息来检测在背景中伪装的自由移动的鱼。我们在两个水下视频数据集（即Fish4Knowledge存储库中的LifeCLEF 2015基准）和西澳大利亚大学（UWA）收集的数据集上评估了该系统。在这两个数据集上，我们实现的鱼类检测F分数分别为95.47％和91.2％，而鱼类物种分类的准确度分别为91.64％和79.8％。据我们所知，这些是这些数据集上报告得最好的结果，