Convolutional neural network-based image recognition of forage seeds

The traditional method of seed identification based on grayscale image processing reduces image data complexity and prediction accuracy. This study improved this method to a more efficient method named convolutional neural networks combined with normalized canonical discriminant analysis (CNN-nCDA), which combines two techniques, nCDA and CNN. CNN-nCDA employs the deep learning framework TensorFlow to distinguish forage seeds based on multispectral imaging, and its seed recognition algorithm is conducted by CNN. The results showed that for seed identification with similar morphology, the accuracy rate of the CNN-nCDA strategy could reach 100.0%. This is a vast improvement over traditional grayscale image processing (62.11%～72.5%), nCDA (90.0%～100.0%), linear discrimination analysis (97.3%～100.0%), and support vector machine (92.4%～97.5%). In conclusion, the CNN-nCDA strategy has high calibration and verification capabilities and holds promise for the purpose of rapid seed identification.