基于卷积神经网络的多层级目标检测方法  

Multi-layer Object Detection Algorithm based on Convolutional Neural Network

作　　者：吴亚熙 岑峰[1] WU Ya-xi;CEN Feng(Tongji University,Institute of electronics and Information Engineering,Shanghai 201804)

机构地区：[1]同济大学电子与信息工程学院,上海201804

出　　处：《软件》2018年第4期164-169,共6页Software

摘　　要：针对目标检测问题中小物体目标普遍难以检测的问题,本文提出了一种基于卷积神经网络的多层级目标检测方法。该算法首先通过对原始图像进行切分,增加对局部图像的关注,然后对原始图片和各个子部分利用深度卷积网络进行检测。对于子层级的检测结果,本文使用子层级抑制筛选算法,该算法主要分为两个步骤,依次是子层级抑制和子层级筛选。子层级抑制阶段的目的是抑制子层级中重叠、截断的目标框,子层级筛选阶段目的是筛选出在主层级中未被检测到的小尺寸目标框。最后输出主层级检测结果和子层级的检测结果,提升检测效果。通过分析在VOC2007数据集上的实验结果,本文提出的多层级目标检测方法对比SSD的检测效果有所提高,特别在小尺寸物体的检测上如bottle,pottedplant类上有较大的提升。Aiming at the small object detection problem, we present a multi-layer object detection method based on convolutional neural network. Firstly, we divide the original image into 4 parts, and then the original image and each sub-part are detected by deep convolution network. For the sub-level detection results, we use the sub-level suppression filtering algorithm, which is divided into two steps, namely, sub-level suppression and sub-level filtering. The purpose of the sub-level suppression is to suppress the overlapping and truncated bounding boxes in the sub-level result. The sub-level filtering is designed to choose small size object bounding boxes that are not detected in the main-level result. Finally, the detection results of the main-level and sub-level detection results are output to improve the detection performance. The Experiment shows that object detection algorithm proposed in this paper is better than the base convolutional neural network model, especially in the detection of small objects such as bottle and plant.

关 键 词：人工智能 目标检测 卷积神经网络 多层级结构 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]