Fig. 2

The flow diagram of the Panicle-SEG algorithm. A Original field rice image. B Mask image. C SLIC superpixel segmentation result. D Automatic labeling. E Training set and validation set building: The patches were augmented and divided into the training set and validation set. F CNN network. G Panicle-SEG-CNN model generation. H Testing sample. I Testing patches generation. J The pre-trained Panicle-SEG-CNN model generated in off-line training is utilized in testing patches classification, and the testing patches were categorized into candidate panicle and confirmed background. K The candidate panicle patches were merged into one image, called the coarse segmentation results. L Entropy rate superpixel image. M Optimized segmentation result. N The final segmentation result was obtained after removing small regions