Manual seed quality assessment in Indian agriculture remains labor-intensive, error-prone, and unscalable, contributing to 20-30% crop losses from poor germination. This paper introduces a novel seed image dataset comprising 5,000+ high-resolution images of paddy and millet seeds across three quality classes (good, medium, bad) collected from Tamil Nadu farms under varying humidity conditions. We present a comprehensive preprocessing pipeline using OpenCV for noise reduction, CLAHE enhancement, GrabCut background removal, and Albumentations-based augmentation (rotation, flipping, brightness adjustment), achieving 95% clean images and

expanding the dataset to 15,000 samples. The pipeline delivers PSNR >30dB and 98% valid segmentation masks, with baseline CNN validation showing 92% accuracy on preprocessed vs. 85% on raw data. Released publicly on Kaggle, this regionally-specific, Non-IID dataset addresses gaps in existing maize/soybean collections by focusing on tropical Indian crops. The work enables deep learning applications for automated seed sorting and supports precision agriculture for smallholder farmers