Deep learning optimization method for counting overlapping rice seeds

Abstract

Abstract Overlapping seeds counting method is an important direction to improve the detec- tion efficiency of rice thousand kernel weight, but the existing image processing methods may misjudge the overlapping distribution when the thin-layer seeds are piled up. In this study, we proposed a deep learning optimization method based on contour grouping pre-labeling for counting overlapping rice seeds. First, the con- tour grouping method—based on the Euclidean distance and divergence function as comprehensive criteria—pre-labels the rice seed contours to reduce the number of potential false classification edge points. The purpose of integrating the pre- labeling results in its feature extraction layer and using a linear combination of inter-class and intra-class error functions as the total is to improve the accuracy of deep learning. Second, based on the improved Faster region-convolutional neural network method, the overlapping rice seeds were accurately counted. An orthogo- nal test is proposed to optimize the full connection layer's parameter configuration, aimed at improving the learning speed. The experimental results demonstrate that the average error rate of rice seeds in a single image (960 960 pixels) is 1.06% and the average recognition time of counting (≤400 seeds) was 0.45 s. The pro- posed method can be used as an effective tool for counting overlapping rice or other crop seeds.

Tasks

Reproductions