Aiming at the problem of low underwater recognition accuracy caused by dense and fuzzy targets in underwater target detection, an underwater target detection algorithm combining attention mechanism and downsampling is proposed.
For the experimental evaluation, the Detecting Underwater Objects (DUO) dataset was used, The dataset was de-duplicated using a perceptual hash algorithm, of which 6671 sheets were used for training, 1111 sheets for testing, and 1111 sheets for validation. There are only a few similar images in the new dataset. Additionally, the DUO dataset includes some images from various underwater scene datasets.The total number of targets in the dataset is 74,515, of which the number of holothurians, echinus, scallops, and starfish are 7,887, 50,156, 1,924, and 14,548, respectively.
Image tag has been converted from json file to txt file, which is convenient for yolo algorithm.
