基于改进DeepLabV3+的荞麦苗期无人机遥感

针对DeepLabV3+语义分割模型计算复杂度高、内存消耗大、难以在计算力有限的移动平台上部署等问题，提出一种改进的轻量化DeepLabV3+深度学习语义分割算法，用于实现无人机荞麦苗期图像的分割与识别。该算法采用RepVGG（Re-parameterization visual geometry group）与MobileViT（Mobile vision transformer）模块融合的方式建立主干网络实现特征提取；同时，在RepVGG网络结构中引入SENet（Squeeze-and-excitation networks）注意力机制，通过利用通道间的相关性，捕获更多的全局语义信息，保证荞麦分割的性能。实验结果表明，与FCN（Fully convolutional networks）、PSPNet（Pyramid scene parsing network）、DenseASPP（Dense atrous spatial pyramid pooling）、DeepLabV3、DeepLabV3+模型相比，本文提出的改进算法在较大程度上降低了模型参数规模，更适合在移动端部署，自建荞麦苗期分割数据集上的语义分割平均像素准确率（Mean pixel accuracy，mPA）和平均交并比（Mean intersection over union，mIoU）分别为97.02%和91.45%，总体参数量、浮点运算次数（Floating-point operations，FLOPs）和推理速度分别为9.01×106、8.215×1010、37.83f/s，综合表现最优。在全尺寸图像分割中，训练模型对不同飞行高度的荞麦苗期分割的mPA和mIoU均能满足要求，也具有较好的分割能力和推理速度，该算法可为后期荞麦补种、施肥养护和长势监测等提供重要技术支持，进而促进小杂粮产业智能化发展。

Segmentation of Buckwheat by UAV Based on Improved Lightweight DeepLabV3+ at Seedling Stage

In view of the problems of high computational complexity, large memory consumption, and difficulty in deployment on mobile platforms with limited computing power in DeepLabV3+ segmentation model, an improved lightweight DeepLabV3+ algorithm was proposed to realize the segmentation and recognition of buckwheat by UAV at seedling stage. The algorithm adopted the fusion of re-parameterization visual geometry group (RepVGG) and mobile vision transformer (MobileViT) modules to establish the backbone network for feature extraction. At the same time, the squeeze-and-excitation networks (SENet) attention mechanism was introduced into the RepVGG network structure to capture more global semantic information by using the correlation between channels, and ensure the performance of buckwheat segmentation. Experimental results showed that compared with fully convolutional networks (FCN), pyramid scene parsing network (PSPNet), dense atrous spatial pyramid pooling (DenseASPP), DeepLabV3, and DeepLabV3+ models, the improved algorithm proposed greatly reduced the model parameters, making it more suitable for deployment on mobile terminals. The mean pixel accuracy (mPA) and mean intersection over union (mIoU) on the self built buckwheat segmentation dataset were 97.02% and 91.45%, the overall parameters, floating point operations (FLOPs) and inference speed were 9.01×106, 8.215×1010 and 37.83 f/s, respectively, with the best performance. In the full-size image segmentation, the mPA and mIoU for buckwheat segmentation can meet the requirements at different flight heights, which had good segmentation ability and inference speed. The algorithm can provide technical support for the later buckwheat seed replacement, fertilization maintenance, and growth monitoring, and promote the intelligent development of small and coarse grain industry.