Abstract: The accurate segmentation of watermelon fruit in UAV(unmanned aerial vehicle) image is the premise of watermelon counting and yield estimation. This paper proposed a segmentation model of watermelon fruit based on an improved UNet network to address the problems of false segmentation and inaccurate detail edge segmentation of UAV watermelon images due to complex field background, uneven illumination, and insignificant features. The visible light image of the UAV in the early ripening stage of watermelon was collected to construct the semantic segmentation dataset of watermelon fruit. An efficient channel attention mechanism was introduced in the downsampling process to enhance the feature weight of the fruit region, and a dual attention mechanism was added in the skip connection part to establish rich context dependency based on local features, so as to improve the feature extraction ability of the target region. Then, the feature map and class activation map were used to visually explain the prediction process of the model. Experimental results showed that the Accuracy, Precision, Recall, F1Score and Intersection over Union(IoU) of the model were 99.03%, 92.67%, 90.55%, 91.21% and 84.71%, respectively, and the processing time of an individual image was 0.145 s. This model can effectively capture the fruit features in the UAV watermelon image in the early maturity stage, accurately identify the fruit regions with complex background under natural environment, and has good segmentation effect and generalization ability. It can provide theoretical basis and technical support for the use of UAV remote sensing technology to count the number of watermelon in the field and estimate the yield at the early maturity stage.

Key words: watermelon fruit, deep learning, UAV image, semantic segmentation, attention mechanism

摘要： 无人机图像中的西瓜果实精准分割是进行西瓜计数和产量预估的前提。针对无人机西瓜图像因存在田间背景复杂、光照不均匀、特征不显著等情况容易导致误分割和细节边缘分割不精确的问题，提出一种改进UNet网络的西瓜果实分割模型。首先采集西瓜成熟前期的无人机可见光图像，构建西瓜果实语义分割数据集；其次在下采样阶段引入高效通道注意力机制，增强果实区域的特征权重，并在跳跃连接部分增加双注意力机制，基于局部特征建立丰富的上下文依赖关系，提高对目标区域的特征提取能力；最后使用特征图和类别激活映射图对模型预测过程进行可视化解释。结果表明，该模型的准确率、精确率、召回率、F1Score值和交并比分别为99.03%、92.67%、90.55%、91.21%和84.71%，单幅图像分割时间为0.145 s。该模型能够有效捕获成熟前期的无人机西瓜图像中的果实特征，准确识别自然环境中复杂背景的果实区域，具有良好的分割效果和泛化能力。为利用无人机遥感技术统计大田西瓜数量和成熟前期产量预估提供理论依据和技术支持。

关键词: 西瓜果实, 深度学习, 无人机图像, 语义分割, 注意力机制