基于MobileNet的多尺度感受野特征融合算法

Multi-Scale Receptive Field Feature Fusion Algorithm based on MobileNet

针对轻量化网络在目标检测中检测精度低的问题,提出了一种以MobileNet为基础网络的轻量级目标检测网络MobileNet-RFB-ECA。针对目标多尺度特性,采用基于轻量化扩充感受野模块(RFB)的特征金字塔网络结构增强网络对目标多尺度特性的适应性。与此同时,针对复杂注意力模块导致计算量大的问题,在主干特征提取网络添加有效通道注意力机制模块(ECA),提高卷积神经网络的性能。实验结果表明,相较于MobileNet,所提MobileNet-RFB-ECA在PASCAL VOC数据集和KITTI数据集上检测精度分别提高了4.2个百分点和15.4个百分点,模型大小分别为50.3 MB和48.5 MB,平均检测速度为34 frame·s^(-1)。

To address the problem of low target detection accuracy in lightweight networks,a lightweight target detection network MobileNetRFBECA based on MobileNet is proposed.To consider the multiscale characteristics of the target,this study proposes a feature pyramid network structure based on the lightweight extended receptive field block(RFB),which enhances the adaptability of the network to the multiscale characteristics of the target.Moreover,owing to the large computation caused by the complex attention module,an efficient channel attention(ECA)module is added to the backbone feature extraction network to improve the performance of the convolutional neural network.Experiments reveal that compared with conventional MobileNet,the proposed method improves the detection accuracy by 4.2 percentage points and 15.4 percentage points on the PASCAL VOC and KITTI datasets,respectively.In addition,the model sizes of the proposed method are 50.3 and 48.5 MB for the aforementioned datasets,respectively,and the average detection speed achieved is 34 frame/s.