Rice-Variety-Classification-System-Dataset
Description
This dataset was developed as part of a Final Year Project focused on rice grain analysis using a multi-stage ensemble deep learning framework. A custom imaging chamber was designed and constructed to capture rice grain images under controlled and consistent lighting conditions with a uniform background. This ensured high-quality image acquisition suitable for computer vision tasks and reduced variability caused by environmental factors. The dataset creation process began with manually captured images of rice grains placed in trays, where multiple grains were present in each image. These images were manually annotated for instance segmentation to support training of a YOLOv8 segmentation model. The annotated dataset contains images where individual rice grains are separated and labeled, enabling precise object-level learning. A total of 407 high-resolution images were collected for the YOLOv8 segmentation dataset, with each image containing approximately 70 rice grains on average. The trained YOLOv8 model was then used to detect and segment individual grains from raw images. A custom cropping pipeline was applied to extract each detected grain, converting them into individual single-grain images. These cropped images were used to construct the dataset for convolutional neural network (CNN) based classification models, including EfficientNetV2-S, MobileNetV2, and ResNet18. This dataset consists of 9,000 training images (3,000 per class) and 900 validation images (300 per class), organized across three rice variety classes. Finally, for the meta-learning stage, an XGBoost-based ensemble dataset was created using the prediction outputs of the trained CNN models. These predictions were aggregated to form feature vectors, resulting in a tabular dataset consisting of 9,000 training samples and 900 validation samples, again balanced across three classes.
Files
Institutions
- NED University of Engineering and TechnologySindh, Karachi