基于深度学习的水面漂浮物目标检测评估

Evaluation of deep neural networks for floating objects detection

在这篇文章中,我们提出了一个关于水面漂浮物的小型数据集,并分析了几种目标检测模型在数据集上的表现,包括FasterR-CNN,R-FCN和SSD。我们的目的是探究目标检测模型在检测水面漂浮物特别是非物体类别时的特性,并找出权衡精确度和速度后最适合于引导水面清洁无人船的模型。为此,我们制作了一个小型的水面漂浮物数据集,数据集主要包括漂浮水草和漂浮落叶。之后我们通过将预训练模型在水面漂浮物数据集上进行迁移学习,实现了对于水面漂浮物区域的目标检测。我们对比并分析了这些模型的表现,SSD目标检测模型有着更高的精确度,FasterR-CNN模型则能给出更详细的预测,而同时拥有丰富结构特征和相当深度特征的模型对于困难目标有着更好的表现。

In this paper, we present a small dataset focusing on floating objects and analyses the performance of several object detection models when detecting floating objects, including Faster R-CNN, R-FCN, and SSD. We aim to explore the properties of these object detection models when detecting floating objects specifically “stuff” objects and find the speed/accuracy balance for guiding the unmanned surface vessel(USV) to clean rubbish on water surface automatically. To this end, we have created a floating objects dataset which is mainly focused on floating leaves and floating weeds. Then we fine-tune the pre-trained models on our dataset to achieve the ability to detect floating objects. We compare and evaluate the performance of these models, including the average precision on 2 major classes and weighted average precision. Our findings show that the models using SSD as meta-architecture gain higher precision but the models using Faster R-CNN as meta-architecture give more specific predictions, and the models have both rich hierarchies features and deep features perform better when dealing with difficult examples.