A lightweight I2I deep learning method for on-panicle grain in-situ counting and occluded grains restoration

Objective  To address the functional and efficiency limitations of the conventional grain phenotype analysis algorithm of seed analyzers, a deep learning based lightweight general algorithmic framework was designed for two tasks: In-situ counting of on-panicle grains and restoration of occluded grains.

Method  Two complex tasks of on-panicle grains in-situ counting and restoration of occluded grains were decomposed into two stages, and their core stages were modeled as I2I problems. A lightweight network architecture capable of solving the I2I problem was designed based on MobileNet V3, and the data set generation method was designed according to the characteristics of these two tasks. Then the network was trained with appropriate optimization strategies and hyperparameters. After training, the model was deployed and tested with TensorFlow Lite runtime on Raspberry Pi 4B development board.

Result  The algorithm had good accuracy, rapidity and some generalizable performance in the task of on-panicle grain counting. In the task of occluded grains shape restoration, the evaluation accuracy of the restored images in the metrics of area, perimeter, length, width and color score were all over 97%.

Conclusion  The algorithm proposed in this paper can complete the task of on-panicle grain counting and occluded grains restoration effectively, and also has the advantage of being lightweight.