Authors

Abstract

In order to solve the problem of citrus full growth cycle identification in complex scenes, this paper proposed a multi-scale detection model of citrus whole growth cycle in orchard environment. The weighted bi-directional feature pyramid network (BiFPN) is used to combine multiple feature information of high resolution and low- resolution feature layers, and the feature information is extracted by the depth-separable convolution and lightweight New-C3 module. The results show that the average accuracy of the multi-scale detection model proposed in this paper was 91.35%, 92.89%, 94.12%, 90.39% in the young citrus, expanding citrus, ripe citrus and full growth cycle citrus, and the average detection time was 92.60 FPS/s under 1920×1080 image pixels, which meets the real-time detection requirements of citrus orchard.

Abstract in Chinese

为了解决复杂场景下柑橘的全生长周期识别问题，本文提出了一种果园环境下柑橘全生长周期的多尺度检测模型。采用加权双向特征金字塔网络（BiFPN）来融合高分辨率和低分辨率特征层的多项特征信息，且通过深度可分离卷积和轻量型New-C3模块实现特征信息的提取。结果表明，本文提出的多尺度检测模型在生长期、膨果期、成熟期和全生长周期柑橘的平均精度为91.35%，92.89%，94.12%，90.39%，在1920×1080图像像素下的平均检测时间为92.60 FPS/s，满足果园柑橘的实时检测要求。