基于轻量级金字塔网络的种子分选方法研究

针对目前卷积神经网络种子分选方法存在识别精度不高、模型参数量大、推理速度慢且难于部署等问题，提出了基于轻量级金字塔空洞卷积网络的种子分选方法；该网络提出了残差空间金字塔模块，利用不同扩张率的空洞卷积扩大感受野,更有效的提取多尺度特征；再结合深度可分离卷积技术减少模型参数量和计算复杂度；在网络结构中引入轻量级注意力机制模块，利用局部跨通道交互方式关注重要的信息，提高种子关键特征提取能力；实验结果表明，提出网络参数量仅为0.13M，在玉米和红芸豆数据集上准确率高达96.00%和97.38%，在NVIDIA Quadro板卡上识别单张图片时间仅为4.51ms，均优于主流轻量级网络MobileNetv2、Shufflenetv2 和PPLC-Net等，可以满足工业现场实时识别的要求。

Research On Seed Sorting Method Based On Lightweight Pyramidal Network

Abstract: To address the problems of low recognition accuracy, large number of model parameters, slow inference speed and difficult deployment of the current convolutional neural network seed sorting method, a seed sorting method based on lightweight pyramidal dilated convolutional network is proposed. This paper proposes the residual spatial pyramid module, which expands the perceptual field by using the convolution of dilated with different expansion rates, so as to effectively extract multi-scale features. Deeply separable convolution techniques are then used to reduce the number of model parameters and the computational complexity. A lightweight attention mechanism module is introduced into the network structure to improve seed key feature extraction by focusing on important information using local cross-channel interactions. The experimental results show that the proposed network has only 0.13M parametric number, 96.00% and 97.38% accuracy on corn dataset and red kidney bean dataset, and 4.51ms average time to recognize a single image on NVIDIA Quadro boards, which are better than the mainstream lightweight networks MobileNetv2, Shufflenetv2 and PPLC- Net, etc., which can meet the requirements of real-time recognition in industrial sites.