A Data Augmentation Strategy Combining a Modified pix2pix Model and the Copy-Paste Operator for Solid Waste Detection With Remote Sensing Images

Abstract

Solid waste detection is of great significance for environmental protection. In recent years, object detection methods based on deep learning have progressed rapidly. However, it is often extremely difficult to collect sufficient data to train a model with a good performance. In this article, a data augmentation strategy was introduced to generate sufficient synthetic high-quality images for solid waste detection. First, a modified pix2pix model was proposed, in which a local-global discriminator was designed to improve the detailed and global information of the generated images, which are commonly fuzzy with the original pix2pix model. Second, a copy-paste operator was utilized, which simply pastes the bounding box of the generated objects into different images to enhance the diversity of the samples. In this manner, the expanded dataset can be utilized to train different object detection models, for which FPN and Yolo-v4 were introduced as the validation models in this article. The experimental results show that the proposed strategy outperforms the traditional pix2pix method and the generated synthetic images can effectively improve the performance of object detection methods.

Keywords

• Copy-paste
• data augmentation
• local-global discriminator (LGD)
• object detection
• pix2pix