基于改进Faster R-CNN的地面垃圾分类与检测

Classification and detection of ground garbage based on improved Faster R-CNN

为了对常见的地面垃圾类别进行检测,提出改进的Faster R-CNN地面垃圾检测算法。在主干网络部分,选择Resnet-50作为特征提取网络,加入通道注意力机制(SENet),加强对不同通道中有用信息的提取。同时有针对性地提取该网络中的浅层信息,提高地面垃圾中的小目标检测精度,使用特征金字塔结构(FPN)加强特征提取网络多个输出之间的特征融合;在区域建议网络部分,利用数据集的尺度、宽高比例等数据进行统计分析,提出更适合地面垃圾数据集的锚框生成机制;在检测和预测部分,使用RoI Align结构以减少误差,并通过级联结构,分两阶段提高IoU(intersection over union)的阈值,从而改善候选框的精度。试验结果表明:使用改进之后的Faster R-CNN目标检测网络,模型的平均精度均值上升了9.6%,减少了漏检和误检,整体效果较好。

In order to detect the common ground waste,an improved Faster R-CNN ground waste detection algorithm was proposed.In the backbone network,Resnet-50 was selected as the feature extraction network,and the Squeeze-and-Excitation Network was added to strengthen the extraction of useful information in different channels.The shallow information in the network was extracted pertinently to improve the detection accuracy of small targets in ground garbage.The feature pyramid network was used to strengthen the feature fusion between multiple outputs of the feature extraction network.In the part of regional proposals network,through the statistical analysis of the data such as the scale and width height ratio of the data set,an anchor generation mechanism more suitable for the ground garbage data set was proposed.In the part of detection and prediction,the RoI Align structure was used to reduce the error.At the same time,through the cascade structure,the threshold of intersection over union(IoU)was increased in two stages,so as to improved the accuracy of candidate boxes.The experimental results show that by using the proposed algorithm,the model mean average percision increases by 9.6%,and the missed detection and false detection reduce.The overall effect is good.