轻量化SSD目标检测方法研究

Research on Lightweight SSD Target Detection Method

为了提高目标检测在嵌入式或移动端设备运行的可能性,基于SSD(Single Shot MultiBox Detector)框架,结合轻量化神经网络,构建一种轻量化SSD目标检测模型,称其为快速且精准的SSD(Fast and Accurate Single Shot Detector,FA-SSD).该方法采用轻量化卷积神经网络ESPNet作为基础网络,使用反卷积模块融合深浅层特征信息,并做轻量化处理,均衡模型尺寸和检测精度.实验结果表明,该方法相比原经典SSD算法具有更少的网络参数量和计算复杂度,在参数量上减少了47.3%,每秒处理图像帧数比经典SSD算法提升3.7倍.在VOC2007数据集中的测试平均精度均值(mAP)结果可以达到73.6%,和经典算法的结果相差无几,从而在保证检测精度的同时提高检测速度.

In order to improve the possibility of target detection running on embedded or mobile devices,we constructed a lightweight SSD target detection model called Fast and Accurate Single Shot Detector(FA-SSD),based on the SSD(Single Shot MultiBox Detector)algorithm framework as well as lightweight neural network design ideas.Taking a lightweight convolutional neural network ESPNet as the basic network,the FA-SSD method uses a deconvolution module to fuse deep and shallow feature information and performs lightweight processing to balance the model size and detection accuracy.The experimental results show that the proposed method has fewer network parameters and less computational complexity in comparison with classical SSD algorithm,which still has a faster prediction speed,It reduces the amount of parameters by 47.3%,and the number of image frames processed per second is 3.7 times higher than that of the classic SSD algorithm.The test mean average precision(mAP)result on the VOC2007 data set is 73.6%.The results prove that the proposed algorithm can improve the detection speed with a high precision ensured.