LEPD-Net: A Lightweight Efficient Network with Pyramid Dilated Convolution for Seed Sorting

Abstract

To achieve long-term economic growth, competitiveness, and sustainability, speed and accuracy are the key requirements when it comes to seed purity sorting. However, current seed sorting methods suffer from large number of model parameters and computational complexity, make it a great challenge to deploy them in real-time applications, especially on devices with limited resources. To issue above problems, in this paper, a lightweight efficient network with pyramid dilated convolution, namely LEPD-Net, is proposed for seed sorting. First, a residual spatial pyramid module (RSPM) is elaborately designed, which uses dilated convolution with different dilation rates to enlarge the structural characteristics of the receptive field and effectively extracts multi-scale features. Then the depth-wise separable convolution to reduce the amount of model parameters and the computational complexity. In addition, to further improve the performance, a novel lightweight coordinate attention module is introduced, which uses the local cross-channel interaction to obtain the attention value of each channel and strengthen the network's ability to learn seed key features. Finally, the seed sorting task is completed through the learned features. Experimental results show that our proposed method achieves an accuracy of 96.00% and 97.25% on the Maize dataset and Red Kidney Bean dataset, respectively. The number of parameters is only 0.26M, which is far less than state-of-the-art networks (e.g., MobileNetv2, Shufflenetv2, and PPLC-Net).