Diffusion model-based image generative method for quality monitoring of direct grain harvesting

Abstract

In direct grain harvesting, precisely detecting grain breakage and impurity rates is crucial for optimizing harvesting efficiency and enhancing grain quality. Deep learning models are increasingly employed in quality monitoring for direct grain harvesting, facilitating accurate time detection of breakage and impurity rates during the harvesting process through image processing. However, deep learning models are often constrained by the grain dataset, which is limited by difficulties in collecting, poor diversity, and few image samples of breakage and impurities, which increases the likelihood of misrecognition and omissions. Therefore, we proposed an image generation method based on the Denoising Diffusion Probabilistic Model (DDPM), which incorporates a designed spatial and channel attention block (SCA-Block) into the U-Net architecture within the DDPM, resulting in the Dual Attention Diffusion Model (DADM). The results of the generated experiments showed that DADM outperformed DDPM and Generative Adversarial Networks (GANs) in terms of FID scores for the corn, rice, and soybean datasets with 45.12, 50.7, and 36.68 respectively. The results of the segmentation experiments showed that DADM enhancement can effectively improve the performance of segmentation models compared to traditional enhancement. To verify the scalability of DADM, we also conducted experiments on the pest and disease detection dataset, and the results showed that the DADM enhancement still worked significantly, with a 5.07% improvement in the MIoU. Our research provides new insights into the enhancement of agricultural image datasets and, through image generation, solves the problem of difficult agricultural image collection due to factors such as season, climate, and environment, and then improves the training effect of downstream tasks such as semantic segmentation and target detection, promoting the comprehensive application of deep learning in agriculture.