Improving Plastic Bottles Recognition with Data Augmentation Methods

Plastic bottles on the surface of water are a major source of water pollution, severely damaging aquatic ecosystems. To achieve effective cleaning of such pollution, surface cleaning robots can be used for automated recognition and salvage. However, existing surface target recognition methods are limited by insufficient coverage of dataset samples, resulting in poor adaptability and limited recognition accuracy. To address this issue, for the original image dataset, this paper employs six data augmentation methods, including flipping transformation, translation transformation, homomorphic filtering, super-resolution, histogram equalization, and Gaussian noise, to independently generate augmented image datasets. For each augmented dataset, Faster R-CNN and YOLOv5 algorithms are used for multi-target and single-target recognition training, respectively. Experiments show that the performance of both algorithms after data augmentation is significantly better than that of the original dataset in both single/multi-target recognition tasks. Specifically, the combination augmentation strategy, i.e., combining all images, achieves the most prominent improvement, and YOLOv5 achieves a recognition accuracy of 100% at different confidence thresholds in single-target recognition tasks on the combined augmentation dataset. The study indicates that data augmentation methods can effectively improve performance and provide a reliable solution for target recognition.