面向多目标图像检测的轻量级SSD网络

Lightweight SSD Network for Multi-target Image Detection

伴随着更低成本和更高性能的工业需求趋势,如何解决神经网络模型中的计算复杂度过高问题就显得尤为重要。为降低原始SSD图像目标检测与识别算法的复杂度,该文提出了一种改进SSD的轻量级网络模型SL-SSD。首先,使用轻量级卷积神经网络ShuffleNet替换了原始SSD网络中的VGG-16模块,降低了网络的计算复杂度;其次,在原始SSD网络的附加特征提取模块上,并行使用小、中、大三种尺度的空洞卷积特征提取方式,实现了网络参数共享、降低了网络复杂度,提升了对小目标的检测精度。最后,在PASCAL VOC07+12标准图像数据集上进行数据实验。结果表明,SL-SSD网络模型在Nvidia GeForce RTX 1080 GPU上取得了70.5%的平均检测精度,达到134MFLOPs计算力。相比于原始的SSD算法以及大部分目标检测算法,改进的算法较好地平衡了检测与识别精度及计算力之间的关系。

With lower cost and higher performance of the industrial demand trend,it is particularly important to solve the problem of high computational complexity in neural network models.In order to reduce the complexity of the SSD image detection,a lightweight network model SL-SSD based on SSD is proposed in this paper.Firstly,a lightweight convolutional neural network ShuffleNet is used to replace the VGG-16 in the SSD network,thus reducing the computational complexity of the network.Secondly,on the additional feature extraction module of the SSD network,the small,medium and large scale feature extraction methods of hollow convolution are used in parallel to realize the network parameter sharing,the network complexity is reduced,and the detection accuracy of small targets is improved.Finally,the data experiment is carried out on PASCAL VOC07+12 standard image data set.The results show that the average detection accuracy of SL-SSD on Nvidia GeForce RTX 1080 GPU is 70.5%,and the computing power is 134MFLOP.Compared with the SSD algorithm and most of the target detection algorithms,the improved algorithm balances the relationship between accuracy of detection and computational force.