基于SSD和MobileNet网络的目标检测方法的研究

Research on Target Detection Method Based on SSD and MobileNet Network

为了提高计算机视觉中目标检测的一种基本模型SSD在多任务场景中的准确率和效率,基于深度学习的相关理论研究,结合一种轻量级的深层神经网络MobileNet的基本思想,构建了一种结合特征金字塔的多尺度卷积神经网络结构。利用Tensorflow平台完成了以下一些工作:第一,对低层卷积层的特征图进行区域放大,保留更多的目标特征信息,再对高特征层进行特征提取;第二,在对重叠目标候选区域进行过滤的时候,基于非极大值抑制的方法和思想设置阈值消除冗余的目标候选区域,使得产生的负样本的数目减少,使模型效果逐步趋于稳定;第三,针对目标检测中的预测区域与真实区域在匹配过程中所产生的正负样本进行处理,用于保证模型的稳定性等。基于以上方法研究,使得模型对多目标识别的速度有所加快,鲁棒性更好,准确率更高,同时也适当降低了对硬件配置资源的需求。

In order to improve the accuracy and efficiency of the SSD(single shot multibox detector)in the multitask scene,which is a basic model of target detection in computer vision,according to the deep learning and relevant theoretical research,this paper constructs a multi-scale convolutional neural network structure with feature pyramid.It contains the basic ideas of a lightweight deep neural network called MobileNet.The following work is completed on Tensorflow platform.First,it amplifies the feature map of the low-layer convolution layer to reserve more target feature information,and then extracts the feature information of the high layer.Second,while filtering the overlap target candidate areas,it sets the threshold value to eliminate redundant target candidate areas based on the idea of non-maximum suppression.It reduces the number of negative samples and makes the model’s effect become stable gradually.Third,to ensure the stability of the model in this paper,it conducts the positive and negative samples in the target detection process of matching between the predicted region and the real region.Based on the research methods above,the recognition speed of the model in multi-target tasks has been accelerated,the robustness is better,the accuracy is higher,and the demands of the hardware configuration resources are reduced appropriately at the same time.