基于改进扩散模型的小样本变电缺陷图像生成方法OA

Due to the high severity and low frequency of substation defects in expander topping and the small number of defect image samples,it is urgent to expand the small-sample defect images through image generation methods to support the effective improvement of substation defect image recognition algorithms.This paper proposes a small sample genera-tion method for transformer defect images based on the improved diffusion model.Firstly,aiming at the problem that the generative model trained with small samples is difficult to control the authenticity of defects,an improved generative dif-fusion model is proposed.By adding the ResNext lightweight control branch,the precise generation of diverse defect details is achieved.Then,aiming at the problem of background distortion when generating large images and reducing the learning difficulty of the model and the video memory resources for generating large images,this paper proposes to gen-erate through defect mask regions and fuse the defect regions into the original image.Finally,in view of the lack of a quality evaluation method for defect image generation,this paper comprehensively evaluates the quality of the generated defect images by integrating image quality indicators,including Frechet inception distance(FID),inception score(IS),structural similarity index measure(SSIM),visual information fidelity(VIF),and the improvement index of the target de-tection model(including F1 score).Experiments show that,compared with the GAN and SD series generative algorithms,the method proposed in this paper can be adopted to improve the image quality index of small-sample transformer defect image generation by 25％.Compared with the object detection algorithm trained only with real defects and normal sam-ples,after adding normal images and the defect images generated using the method proposed in this paper,the F1 score of the trained object detection algorithm is increased by more than 12％on average.