一种基于深度学习的新型小目标检测方法

A NEW METHOD OF SMALL TARGET DETECTION BASED ON DEEP-LEARNING

快速、精准的目标检测方法是计算机视觉领域的研究热点之一,目前通用的目标检测模型主要包括两个部分,候选区域提取和分类器设计。基于卷积神经网络CNN和超像素算法提出了一种新型面向微小目标的检测方法。首先对目标图像进行超像素过分割,然后提取过分割区域的特征并以此进行区域合并,最后提取候选区域。与传统建议区域提取方法相比,本方法能够在保证召回率的前提下大量减少候选区域的数量。为了克服小目标特征提取的困难,本算法利用多尺度和多层次CNN提取候选区域的中高层语义信息进行目标分类。通过对车辆年检标示数据的实验表明提出的基于超像素过分割候选区域提取算法具有较高的召回率,在同等候选区域数量的情况下与Edge Box、Bing、Selective search等方法相比分别提高2%、2.4%和3.5%,同时基于多层次多尺度的目标分类算法能有效降低误检率,提高检测率。

Accurate and fast object detection is one of the research topics in computer vision. At present,the general target detection model mainly consists of two parts,the extraction of candidate regions and the design of classifier. This paper innovatively proposes to apply convolutional neural network（ CNN） and super pixel to the detection of a new small target. Firstly,we employed SLIC algorithm to over-segment the image. Then,we extracted the features of the over segmentation region and merged the regions. Finally,candidate regions were extracted. Compared with the traditional proposed region extraction method,our proposed method reduced the number of candidate regions on the premise of ensuring recall rate. To overcome the difficulty of feature extraction of small targets,our algorithm used multi-level and multi-layer CNN to extract semantic information of the middle and high level of candidate regions for target classification.Experiment on detecting vehicle inspection mark shows that our method achieves better recall rate（ increased by 2%,2. 4%,3. 5%） compared with the state-of-the-art method including Bing,Selective search,and Edge box. Meanwhile,the multi-level and multi-scale target classification algorithm can effectively reduce the false detection rate and improve the detection rate.