针对安全帽检测中,目标形状、尺度变化大,易出现漏检、误检等问题,提出了一种基于改进级联基于区域的卷积神经网络(CascadeR-CNN)的安全帽检测算法.首先,对ResNet50进行改进形成D-ResNet50,利用可变形卷积仅增加少量参数就可增大感受野的特性,对特征提取网络的C2～C5卷积层进行重塑,提高网络对目标几何变换的适应能力和特征提取能力.其次,将D-ResNet50作为主干网络引入Cascade R-CNN,形成级联目标检测器,在每个阶段对正负样本重采样,抑制误检问题.再次,对递归特征金字塔进行改进,更高效地进行多尺度特征融合,并且基于反馈信息对特征进行二次处理,增强特征表达,提高网络的分类和定位能力.最后,使用Soft-非极大值抑制(Soft-NMS)进行后处理,进一步解决漏检问题.提出的方法在Hard hat workers数据集上的AP值相比检测基线提高了3.5％,与 SparseR-CNN、TridentNet、VFnet 等先进算法相比分别提升了4.7％、5.9％、2.3％等.
Abstract
Aiming at the problems of large changes in target shape and scale,which are prone to missed detection and false detection in safety helmet detection,a safety helmet detection algorithm based on improved Cascade Region-based Convolutional Neural Networks(Cascade R-CNN)is proposed.Firstly,ResNet50 is improved to form D-ResNet50,which can increase the perceptual field by using the feature of deformable convolution with only a small increase of parameters,and reshape the C2 to C5 convolutional layers of the feature extraction network to improve the network's adaptability to target geometric transformation and feature extraction capability.Secondly,D-ResNet50 is introduced into Cascade R-CNN as the backbone network to form a cascade target detector.Convolutional Neural Networks(CNN)to form a cascade target detector,resampling positive and negative samples at each stage to suppress the false detection problem.Thirdly,the recursive feature pyramid is improved to perform multi-scale feature fusion more efficiently,and the features are processed twice based on feedback information to enhance feature representation and improve the classification and localization ability of the network.Finally,post-processing is performed using Soft-Non-Maximum Suppression(Soft-NMS)to further solve the problem of missed detection.The proposed method improves the AP value on Hard hat workers dataset by 3.5％compared to the detection baseline,and by 4.7％,5.9％,2.3％,etc.compared to the advanced algorithms such as Sparse R-CNN,TridentNet,and VFnet,respectively.
维普
万方数据