基于改进YOLOv4模型的自然环境下梨果实识别

Recognition of pear fruit under natural environment using an improved YOLOv4 model

针对自然环境下梨果实识别场景中存在梨果实颜色与背景颜色相近、遮挡、重叠等因素导致的识别困难的问题,本试验提出1种基于改进的YOLOv4网络模型梨果实识别的方法,使用的神经网络模型以CSPDarknet53作为主干特征提取网络,将空间金字塔池化(SPP)中的最大池化法改为平均池化法,以适应目标与背景颜色相近的场景,更多地保留目标信息;将SPP模块前后的卷积、PANet中的部分卷积以及输出部分的卷积替换为深度可分离卷积,在保证卷积效果不变的效果下减少网络模型所占空间。使用训练后的改进YOLOv4模型对未参与训练的图像样本进行测试,改进后的模型所占空间比原模型下降44%,召回率达到85.56%,比原模型提高了1.29%,mAP达到90.18%,比原模型提高了0.1%。实验结果表明,本文算法对自然环境下近色背景的梨果实的识别具有良好的查全率与精确率,能够较好地对梨果实进行识别,可为实现梨果园的自动采摘和产量预测提供技术支持。

Aiming at the recognition problems of pear fruit under natural environment including similar background color, covering and overlaping,this paper improved YOLOv4 network model for pear fruit recognition. The neural network model was applied with CSPDarknet53 as the backbone feature extraction network. In order to retain more target information under similar colors background as target, the maximum pooling method was replaced by the average pooling method in spatial pyramid pooling(SPP). The convolutions were replaced by depth-wise separable convolutions in the partial convolutions in PANet, output part, before and after the SPP module to reduce the space occupied by network model. The improved model after training was used to test the image samples that did not participate in the training, and the improved model took 4 % less space with 85.56% recall rate that was improved by 1.29%. And the m AP reached 90.18%, which was 0.1% higher than the original model. The experimental results showed that the algorithm in this paper had a good recall and precision rate for the recognition of pear fruits with near-color background under natural environment. The algorithm can recognize pear fruits well, providing a new solution for automatic picking and yield forecast of pear orchards.