Real-time fish localization with binarized normed gradients

Fast and accurate fish localization is an important step for fish detection, identification, counting and tracking. In this paper, we introduce how to localize the fish with an efficient way, which can capture almost all fish locations in an image. First, we exploit the normed gradients (NG) feature of 8×8 image windows to discriminate the fish from the background, and then we binarize the NG feature to accelerate the fish localization. As there is no existing appropriate dataset, we make a dataset of underwater imagery to achieve fish localization. The dataset contains 9,963 images of underwater videos for training, validation and testing. The details about how to label the fish of this dataset further be showed. Last, we evaluate our method on this dataset. Experiments show that our method is fast and efficient, and fish localization takes only about 0.00234 sec. per image (400 fps on an Intel i5-3540 CPU) and achieves 97.1% recall with 1000 proposals. This method satisfies computational efficiency and high detection rate simultaneously.