Similar target replacement for remote sensing object detection data augmentation

Nowadays remote sensing image object detection algorithms are highly relied on the development of deep learning technology. Data augmentation to dataset images is an important way to enhance model's generalization ability. Current augmentation methods of remote sensing object detection algorithms still use general object detection methods. There is an urge need to develop methods that focus on the properties of remote sensing objects. This paper raised a data augmentation method named Similar Targets Replacement (STR) that based on remote sensing images and designed the STR process. First, the sample library is built to collect statistics of samples' categories. Second, categories of the dataset is divided into several similar target categories. Then, to solve the problem of disbalanced sample amounts, minority sample compensation is designed to balance the proportion of samples from different categories by controlling the probability of input samples. Finally, sample replacing preprocess mechanics is designed by using appropriate transforms for each different categories as the preprocess methods. Experiments on DOTA dataset shows the mAP of DCL detection algorithm using STR augmentation raises 1. 34 compared to baseline. Model's detection accuracy to similar categories raises and ability to categories with fewer number of samples is strengthened.