Seedling counting and localization algorithm based on semi-supervised learning

Accurately measuring number and location of rice plants plays a vital role in rice breeding and cultivation．Traditional field sampling surveys are time-consuming, labor-intensive, and inaccurate．UAV remote sensing technology provides an effective way to intelligently and accurately count and locate rice seedlings．Aiming at a problem of UAV image processing required a large amount of data annotation, a seedling counting and localization algorithm based on semi-supervised learning was proposed, called sparse point annotation network（SPANet）．SPANet was built according to Teacher-Student training paradigm, and with a view to problem of inaccurate generation pseudo-labels from sparse point annotations, the proposed point aggregation module（PAM）could generate high-quality pseudo-labels from limited point annotation information．To verify performance of SPANet, a rice counting and localization dataset containing 300 high-resolution（1600 pixel ×1600 pixel）images was constructed, and conducted detailed experiments at annotation ratios of 50% and 80%．Results showed that when annotation ratio was 80%, mean absolute error（MAE）and mean square error（RMSE）of SPANet were 15.7 and 20.2 respectively, thus achieving accurate and efficient counting and localization of rice seedlings．