面向目标检测的SSD网络轻量化设计研究

在基于深度卷积神经网络的目标检测方法中,模型的参数量动辄数十兆字节,在计算资源有限的移动终端等边缘设备中部署这样的大模型比较困难。为了解决这个问题,本文在Single Shot MultiBox Detector(SSD)的基础上联合轻量化网络设计和参数量化两种技术来实现网络模型的轻量化。首先,基于ResNet50和MobileNet我们重新设计了SSD目标检测框架,并训练了一个全精度参数模型。然后,在全精度参数模型的基础上,采取逐块量化的策略将特征提取层中卷积层的参数精度降低到三值(零和正负一)。实验结果表明,本文提出的联合方案在Pascal VOC2007数据集上测试能够达到72.54%的mAP,和其他业界领先的轻量级目标检测方法相比检测精度更高且能使模型占用的内存空间更小。

A Light-weighted SSD Network Design for Object Detection

In the object detection method based on deep convolutional neural networks, the parameter amount of its model is often tens of megabytes, so it is difficult to deploy such a large model in edge devices such as mobile terminals with limited computing resources. In order to solve this problem, this paper combines lightweight network design and parameter quantification to achieve the lightweight of the network on the basis of one-stage object detector (SSD). First, we redesigned the SSD object detection framework based on ResNet50 and MobileNet and trained a model with full precision weights. Then, based on the full-precision weight model, a block-by-block quantization strategy is adopted to reduce the weight accuracy of the convolutional layer in the feature extraction layer to three values (zero, plus and minus one). The experimental results show that the joint scheme proposed in this paper can test 72.54% mAP on the Pascal VOC2007 dataset. Compared with other industry-leading lightweight object detection methods, the detection accuracy is higher and the model can occupy less memory space.