CitrusYOLO: A Algorithm for Citrus Detection under Orchard Environment Based on YOLOv4

文献类型： 外文期刊

第一作者： Chen, Wenkang

作者： Chen, Wenkang;Lu, Shenglian;Chen, Ming;Li, Guo;Lu, Shenglian;Chen, Ming;Liu, Binghao;Qian, Tingting

作者机构：

关键词： Fruit recognition; Improved YOLOv4; Deep learning; Real-time detection

期刊名称：MULTIMEDIA TOOLS AND APPLICATIONS （ 影响因子：2.577； 五年影响因子：2.396 ）

ISSN： 1380-7501

年卷期：

页码：

收录情况： SCI

摘要： Achieving rapid and accurate detection of tree fruits under natural environments is essential for many precision agriculture application (such as harvesting robots and yield estimation). A real-time citrus recognition method was proposed in this paper by improving the latest YOLOv4 (You Only Look Once version 4) detector for using in orchard environments. The Canopy algorithm and the K-Means + + algorithm were used to automatically select the number and size of prior frames corresponding to the image dataset. Then, an attention mechanism module is added in front of the output layer of each feature of different scales, and a depthwise separable convolution module is added before the upsampling of the network neck to better detect clementines in complex backgrounds. Finally, the network is pruned using the scientific control-based neural network pruning (SCOP) algorithm, and the parameters of the pruned model were fine-tuned to restore some recognition accuracy. Five common used deep learning algorithms including the Faster R-CNN, SSD, YOLOv3, YOLOv4 and Detectron2 were compared to verify the effectiveness of the proposed method. The experimental results showed that our improved YOLOv4 detector works well for detecting different growth periods of citrus in natural orchard environment. The average accuracy increase from 92.89 to 96.15%, the detection time for each image is 0.06s, both are superior the above five algorithms. While the model size was down from 250 MB to 187 MB. This would promise the proposed method being suitable for orchard yield estimation and the development of fruit harvesting robots.

分类号：

  • 相关文献
作者其他论文 更多>>

微信小程序