Abstract: In rice cultivation, weeds are a major factor affecting rice yield. UAVs are increasingly used in the field of smart agriculture. To address issues such as image shaking during image capture and motion blurring when rice weeds are photographed, this paper proposes adding a superresolution module before the segmentation network to solve the problem of unclear images. To improve the accuracy of image segmentation, this paper proposes to convert the image from RGB to Lab color space, so as to increase the differentiation between rice and weeds in computer vision. The Lab values of rice and weeds are weighted as a loss function parameter, integrating more information from the original image to improve the network training accuracy. A local attention mechanism is introduced in U—Net to focus on the important parts of the image, reduce the influence of irrelevant areas, and strengthen the segmentation ability for rice and weeds images, thereby improving the overall performance of the network. The experimental results show that the improved network achieves an Accuracy of 98.1%, Precision of 95.4%, Recall of 96.9%, and mIoU of 84.2%.

Key words: weed in rice fields, neural networks, superresolution, Lab color space, attention mechanism

摘要： 在水稻种植中，杂草是影响水稻产量的重要因素。无人机在智慧农业领域应用日益广泛。针对无人机在图像采集时发生抖动以及稻田杂草拍摄时产生运动模糊的情况，通过在分割网络前增加超分辨率模块来解决图片不清晰的问题；为提高图像分割准确率，提出将图像由RGB转化成Lab颜色空间，从而增加水稻和杂草在计算机视觉上的区分度，同时将水稻与杂草的Lab数值加权作为损失函数参数，融合更多的原图信息，提高网络训练精度；在U—Net中增加局部注意力机制，关注图像中重要的部分，减少无关区域的影响，加强对水稻杂草图像的分割能力，提升网络性能。试验结果表明，改进后网络图像分割的准确率达98.1%，精确率达95.4%，召回率达96.9%，平均交并比mIoU达84.2%。

关键词: 稻田杂草, 神经网络, 超分辨率, Lab颜色空间, 注意力机制