Dreambooth-based Image Generation Methods for Improving the Performance of CNN

With the continuous progress of digital medicine, how to improve the accuracy of Convolutional Neural Network (CNN) algorithm in the absence of sample size has become a difficult problem. This paper explores how to use the stable diffusion image generation model to generate training samples to improve the accuracy of the CNN image recognition model. In this study, the dreambooth algorithm was used to fine-tune the stable diffusion model and learn the features of brain tumour MRI images, and then the stable diffusion model is used to generate images, and the generated images are used to improve the accuracy of CNN image recognition model. This study first used the dreambooth algorithm to train brain MRI images. After debugging the parameters of the dreambooth algorithm, a relatively good version of the weight was selected. Then this study used the same diffusion model to generate images. After debugging the sampler and adjusting the number of training steps, many attempts were made to get a good sample image, so batch generation is performed. After batch selection of the generated samples, the training set is obtained. The images generated by stable diffusion are then added to the database. The experimental results show that when about 800 images are generated for training, the trained CNN recognition accuracy rate is 60%, which is 3% higher than that of a smaller training set.