Plastic bottles on the water surface are important factors causing water source pollution. The low efficiency of traditional manual salvage makes the automated fishing technology of plastic bottles on the water surface extremely urgent. However, the recognition methods of plastic bottles on the water surface generally have the problem of dataset limitations. The limited datasets lead to insufficient extraction of plastic bottle feature information, limited generalization ability of the model, and thus reduced recognition efficiency. Therefore, to address this problem, this paper proposes a data augmentation method based on flip transformation. The 1200 images with plastic bottles in the original dataset are flipped horizontally and vertically, and then 1200 horizontally-flipped images and 1200 vertically – flipped images are generated, which increased the number of images from 1200 to 3600. Experimental results show that after the dataset is processed by flip transformation, the performance of both the Faster R-CNN (Regions with Convolutional Neural Network) and YOLOv5 algorithms has been significantly improved. In the multi-object and single-object recognition tasks of the Faster R-CNN algorithm, the result on the comprehensive dataset with 3600 images perform better than that on the original dataset. Moreover, in the single-object recognition process of the YOLOv5 algorithm, the result on the comprehensive dataset has obvious advantages, with an accuracy reaching \(100\%\) . In conclusion, flip transformation can effectively optimize the performance of object recognition algorithms and significantly improve algorithm’s recognition ability caused by limited datasets.