基于空间信息增强的轻量化玉米果穗品质识别

果穗是籽粒的聚集形态。为实现轻量化卷积神经网络对玉米果穗品质的准确、快速识别，提出了一种结合轻量化主干和轻量化通道池化注意力模块（Lightweight channel pooling attention，LCPA）的玉米果穗品质识别模型LCPA-Ghost。首先，采用Ghost网络实现轻量化处理，减少训练成本和冗余信息，提升模型的特征学习能力。其次，将LCPA模块增加到Ghost模块的捷径连接中，在引入少量参数的情况下，弥补空间信息捕获能力的不足，保证模型识别准确率。实验以正常、籽粒杂乱、霉变、杂色和缺粒果穗为研究对象，采集并制作了包含1571张果穗图像的基础数据集。实验结果表明，LCPA-Ghost模型的测试识别率达98.12%，与CorNet相当，而模型参数量仅为2.40M，单张识别速度为19.08ms，提升9.8%。LCPA-Ghost模型为玉米果穗品质的轻量化识别提供了可行的实验方法。

Lightweight Maize Ear Quality Identification Based on Spatial Information Enhancement

The ear, is the aggregated form of the seeds. In order to realize the accurate and fast recognition of corn ear quality by lightweight convolutional neural network, a corn ear quality recognition model LCPA-Ghost combining lightweight backbone and lightweight channel pooling attention was proposed. Firstly, the Ghost network is used to achieve lightweight processing, reduce training cost and redundant information, and improve the feature learning ability of the model. Secondly, the LCPA module is added to the shortcut connection of Ghost module to make up for the lack of spatial information capture capabilities and ensure the model recognition accuracy by introducing a few parameters. The experiments were conducted with normal, seed disorder, mildew, miscellaneous color and missing grain ears, and a base dataset containing 1571 images of ears was collected and produced. The experimental results show that the test recognition rate of LCPA-Ghost model reaches 98.12%, which is comparable to CorNet, while the number of model parameters is only 2.40M, and the single recognition speed is 19.08ms, an improvement of 9.8%.The LCPA-Ghost model provides a feasible experimental method for the lightweight identification of maize ear quality.