<p>In the field of infrastructure maintenance and management, the practical application of deep learning-based anomaly detection models utilizing images or videos as input is advancing to enhance the efficiency of anomaly detection. However, training a supervised image classification model requires substantial amount of data, which is often unavailable. This lack of sufficient training data frequently limits model performance practical application. Data augmentation is often performed as a method to improve accuracy with limited data. In recent years, technology for image generation based on diffusion models has rapidly advanced, and it has been shown that increasing the amount of training data through data generation using diffusion models can improve model performance. However, only general label generation is typically performed, posing challenges in generating rare anomaly data that exist in the real-world scenarios. This study proposes a new data augmentation method combining geometric pattern mask images and diffusion models to address this gap. By capturing the features of the original image in the unmasked areas and generating the masked regions, new images can be generated while preserving the features of the original labels, facilitating the generation of rare anomaly data. The experimental data uses a dataset of tomato leaf lesion images. The change in model performance when training image classification models with limited data using the proposed method is confirmed experimentally. Experimental results showed up to a 19.50% improvement in accuracy with the proposed data augmentation method. Furthermore, additional experiments demonstrated even greater accuracy improvements when combined with other data augmentation techniques. Notably, as this method does not require text prompts for generation, it holds potential for utility across diverse datasets.</p>