结合注意机制和多尺度卷积的YOLO行人检测

YOLO Pedestrian Detection with Attention Mechanism and Multi Convolution Kernel

为提高行人检测的检测性能,本文结合SqueezeNet、注意力机制、空洞卷积和Inception等结构,提出一种基于改进YOLOv4的行人检测算法.改进YOLO在特征增强部分引入残差连接和结合空洞卷积的注意力模块DCBAM,可以从提取到的特征中选择对目标检测重要的信息.此外,结合SqueezeNet的“squeeze-expand”结构和Inception网络的多尺度卷积思想提出Inception-fire模块用于替代网络中的连续卷积层,通过增加网络的宽度达到提升算法性能的效果,同时减少网络的参数.最后,根据行人检测任务的特点并结合Focal loss对损失函数进行改进,分别对正负样本和难易样本添加权重因子,强调对正样本和难分类样本的训练,从而提高网络的检测能力.改进的YOLO算法在INRIA行人数据集上的检测精度能够达到94.95%,相对原YOLOv4提高4.25%,同时参数量减少了36.35%,检测速度也获得13.54%的提升,在行人检测中能够表现出更优秀的性能.

To improve the pedestrian detection performance,this study proposes a pedestrian detection algorithm based on improved YOLOv4 by combining SqueezeNet,attention mechanism,dilated convolution,and Inception structure.An attention module named D-CBAM is proposed which is combined with dilated convolution.It is introduced to the feature enhancement part to select useful information from the extracted features.The residual connection is also used in this part to enhance feature reusability.In addition,an Inception-fire module is proposed by combining the“squeeze-expand”structure of SqueezeNet and the multi-scale convolution kernel structure of Inception,which replaces the continuous convolution layer in the network.Increasing the width of the network improves the performance of the algorithm and reduces network parameters.According to the characteristics of pedestrian detection and focal loss,the loss function is improved.The detection ability is enhanced through the addition of weights to the positive and negative samples and the hard and easy samples respectively and the strengthening of the training on positive samples and hard samples.The detection accuracy of the improved YOLO algorithm on INRIA person data set can reach 94.95%,which is 4.25%higher than that of YOLOv4.The parameters of the model are reduced by 36.35%,and the detection speed is improved by13.54%.In short,the improved algorithm shows better performance in pedestrian detection than YOLOv4.