Shizuoka University, in collaboration with Yamaha Motor Co., Ltd., has succeeded in researching and developing a new method of expanding generated data using conditional image generation AI for machine learning tasks in fields such as agriculture, which require a lot of effort to prepare and create sufficient, high-quality training data.
To realize smart agriculture, a large amount of training data is required in which specific parts of crop images are labeled. However, image data of crops shows various aspects depending on the conditions and environment, and has diverse domain characteristics. There is also ambiguity in labeling flowers, fruits, nodes, ripeness, degree of disease, etc., making it difficult to create large amounts of high-quality, consistent training data.
The research group developed a new method for expanding generated data using conditional image generation AI. In this method, a large amount of image data is extracted from video data taken by unmanned ground vehicles (UGVs) and other vehicles, and global features are first learned using this large amount of image data. Next, a small amount of training image data is prepared and local features are trained, allowing a large amount of training image data with domain features that meet specified conditions to be mechanically and automatically generated.
As a result, we expanded the generated data from daytime images to a small amount of nighttime images for measuring wine grape growth, and verified the effectiveness of various object detection models and keypoint detection models. As a result, we confirmed that a significant improvement in accuracy of 28.7% was expected for the object (BBox) detection task and 13.7% for the part (keypoint) detection task.
This will drastically reduce the labor hours required to prepare training image data (2,400 images) from 600 hours to about 1 hour. In addition, it can be applied to changes in time, weather, fields, and growth stages, and is expected to accelerate the development of agricultural DX (digital transformation) using AI, such as crop growth estimation and yield forecasting.
Paper information:【Computers and Electronics in Agriculture】D4: Text-guided diffusion model-based domain adaptive data augmentation for vineyardshoot detection