自然环境下基于轻量级SSD的柑橘检测研究

Orange Detection Based on Lightweight SSD in Natural Environment

柑橘采摘机器人常被用来替换工人完成采摘任务。为了采摘机器人在自然环境下能更好地检测到成熟的柑橘果实,一方面,文章针对柑橘的生长特点提出一种轻量级single Shot MultiBox Detector(SSD)柑橘检测模型,首先,为了轻量化模型,将基础网络由原来的VGG-16替换成MobileNetv2,其次,为了解决预测特征层默认框生成的随机性及不稳定性,将K-means算法聚类数据集获得的结果替换默认框的大小,最后,为了减小边界框回归误差,将定位损失函数由原来的Smooth L1修改为GIoU;另一方面,文章自建自然环境下的柑橘数据集并对数据集进行优化处理。实验结果表明,文章改进后的轻量级SSD模型在自建的柑橘数据集上的表现不仅比SSD模型的Average Precision(AP)值高了约8个百分点,而且每秒传输帧数(FPS)达到了37.3,达到了实时检测的要求。

Orange picking robots are often used to replace workers. In order to better detect ripe orange fruits by picking robots in the natural environment, on the one hand, the paper proposes a lightweight single Shot MultiBox Detector(SSD) orange detection model based on the growth characteristics of orange. Firstly, in order to lightweight the model, the feature extraction network was replaced by the original VGG-16 into MobileNetv2. Secondly, in order to solve the randomness and instability of the default box generation in the prediction feature layer, the size of the default box was replaced by the results obtained from the clustering data set of k-means algorithm. Finally, in order to reduce the boundary box regression error, the positioning loss function was changed from Smooth L1 to GIoU;On the other hand, the orange data set under natural environment was constructed and optimized. The experimental results show that the performance of the improved lightweight SSD model on the self-built orange data set is not only about 8 percentage points higher than the Average Precision(AP) value of the SSD model, but also the transmission frame per second(FPS) reaches 37.3, which meets the requirements of real-time detection.