Research on Data Augmentation Methods for Small Sample Image Datasets

In the military field, due to special reasons such as confidentiality and adversarial conditions, there is a shortage of target data and a lack of semantic information, leading to difficulties in equipment recognition. To address this, a considerable number of semantically rich virtual simulation images of equipment can be generated through rendering in digital test technology's simulation physics engines, achieving large sample accumulation. However, the significant differences in texture and lighting between the rendered simulation images and real images cause environment perception models trained on simulation images to be unsuitable for real-world scenarios. This research suggests a realistic style image transmission technique for ship equipment that is based on a diffusion model in order to solve this problem. The semantic analysis of the virtual image sceneries is first obtained through the application of semantic segmentation. Next, the diffusion model is used to create actual images from the semantic analysis. Subsequently, the resulting images undergo super-resolution processes to improve both image quality and size. Finally, a range of assessment criteria are used to examine the experimental results on four datasets in a quantitative and qualitative manner. The suggested method outperforms other transfer methods, as demonstrated by its greater mean Intersection over Union (IoU) scores of 17% and 19%, respectively, and its lower Frechet Inception Distance (FID) scores of 49.34 and 99.32. These findings are based on tests conducted on the virtual road dataset TORCS and the maritime ship dataset Vir-Sea. There is a 5% and 15% increase in Learnt Perceptual Image Patch Similarity (LPIPS) scores. The experimental results show that the proposed method can convert virtual equipment images into actual photos with higher quality, providing a new approach to expand equipment image datasets and aiding in improving the training efficacy of target identification models.