Abstract: Realtime acquisition of grain growth density is helpful to reduce the lag in the detection process of combine harvester feeding amount. In order to obtain grain growth density efficiently and in real time, this paper proposes a realtime detection method of rice growth density based on improved DeepLabV3+ semantic segmentation model. In order to reduce the influence of convolution on feature fusion in the decoder, this study adds the Shuffle Attention module after the convolution module of the encoder part and the backbone network, which reduces the semantic loss. By improving DeepLabV3+, the rice image is semantically segmented and the proportion of rice pixels in the whole image is calculated, and the rice growth density detection model is obtained by fitting with the actual measured rice growth density. The results showed that there was a strong linear relationship between rice pixel ratio and rice growth density per unit area, the determination coefficient R2 was 0.68, and the root mean square error RMSE of the test set was 0.46kg/m2, which could meet the needs of field rice growth density detection.

Key words: image segmentation, rice, density detection, DeepLabV3+

摘要： 实时获取成熟期谷物生长密度有助于减少联合收割机喂入量检测过程中存在的滞后性。为高效、实时地获取谷物生长密度，提出基于改进DeepLabV3+语义分割模型的成熟期水稻生长密度实时检测方法。在DeepLabV3+编码器部分的卷积模块后和主干网络后加入Shuffle Attention模块以减少语义损失。通过改进DeepLabV3+对水稻图像进行语义分割并计算水稻像素在整张图片中的占比，并与实际测得的水稻生长密度进行拟合建立模型得到水稻生长密度检测模型。结果表明，单位面积下水稻像素比与水稻生长密度之间存在较强线性关系，决定系数R2为0.68，测试集的均方根误差RMSE为0.46kg/m2，能实现田间水稻的密度检测。

关键词: 图像分割, 水稻, 密度检测, DeepLabV3+