【作者】 陈涛；

【导师】 鲁红英；

【作者基本信息】 成都理工大学 ， 计算机科学与技术， 2022， 硕士

【摘要】 随着深度学习迅猛发展,目标检测的研究与创新也在不断涌现。在目标检测的重难点方向中,小目标检测问题一直备受学者们关注。小目标检测指对图像中小目标样本进行定位识别,如农业中利用监控设施对害虫、病株进行检测提示;畜牧业中利用无人机对放牧动物的行程线路进行监测控制;建筑业中对工人佩戴安全帽等防护设备的情况进行侦测警告;交通领域中无人驾驶车辆对交通标志进行观测分析等。小目标检测在多个行业中都有涉及,具有重大的应用价值和研究意义。小目标检测难点一方面是图像中小目标携带信息量少、特征不明显以及容易被干扰等;另一方面是网络中卷积层数的增加,导致特征层的小目标信息在不断减弱甚至消散。针对上述难点,本文从数据增强方向提升图像内小目标特征信息,对YOLO v5算法的各个模块进行改进提升网络对小目标特征的维持力度,主要工作内容如下:(1)针对现有公开数据集存在小目标样本过少和样本类别不均衡的问题,通过爬虫程序抓取网络公开图像,进行清洗、标注和审查等处理,获得较丰富的小目标数据集,为实验的准确性和可靠性奠定基础;(2)针对当前数据增强算法对小目标特征提升薄弱,甚至削减小目标特征的问题,提出基于区域的多态过采样数据增强算法(RM-Oversampling)。该算法先提取图像中的目标,用传统增强算法对提取目标进行增强,得到多态过采样数据,接着对图像按照区域划分,将增强数据随机填充到区域中。多态过采样算法能增大目标区域的丰富度,提升小目标的多样性,区域随机填充方案能保证小目标分布均匀,提升图像的空间利用率;(3)针对YOLO v5的Focus算法引起小目标特征属性丢失比例过大的问题,提出基于均值优化的A-Focus算法。该算法利用像素周围区域均值信息,将每个区域像素和区域均值通过映射得到新的特征值。均值信息的融入能减缓背景因素对小目标边缘的干扰,提升输出特征层中的小目标信息量;(4)针对CSPNet结构中的卷积层存在特征层感受野小和特征捕获能力差的问题,提出基于空洞卷积的DCSPNet结构,搭建D-YOLO网络。空洞卷积通过提升感受野区域,使得上层特征层的信息更容易保留到下层特征层,从而保证小目标特征充分传递。实验结果表明D-YOLO算法比YOLO v5算法在小目标检测、复杂背景下小目标检测和密集小目标检测上都有显著提升,在评测指标中准确度提高1.65%,召回率提高0.86%,m AP提高1.67%。更多还原

【Abstract】 With the rapid development of deep learning,the research and innovation of object detection are also emerging.Among the important and difficult directions of object detection,the problem of small object detection has been of great interest to scholars.Small object detection refers to the localization and recognition of small object samples in images,such as the use of monitoring facilities in agriculture to detect and prompt pests and disease plants;the use of drones in animal husbandry to monitor and control the travel routes of grazing animals;the detection and warning of workers wearing protective equipment such as helmets in the construction industry;and the observation and analysis of traffic signs by unmanned vehicles in the transportation field.Small object detection is involved in several industries and has significant application value and research significance.The difficulties of small object detection are,on the one hand,the small amount of information carried by small objects in images,the lack of obvious features,and the ease of interference;on the other hand,the increase in the number of convolutional layers in the network,resulting in the information of small objects in the feature layer is constantly weakening or even dissipating.Given the above difficulties,this paper improves the feature information of small targets in the image from the direction of data enhancement and improves each module of the YOLO v5 algorithm to improve the network’s ability to maintain the features of small targets.The main work is as follows:(1)To solve the problems of too few small object samples and unbalanced sample categories in the existing public dataset,crawl the web public images through a crawler program,clean,annotate,review,and so on,to obtain a richer small object dataset,laying the foundation for the accuracy and reliability of the experiment.(2)To deal with the problem that current data enhancement algorithms are weak in enhancing small object features or even cutting small object features,a region-based multi-style oversampling data enhancement algorithm(RM-Oversampling)is proposed.The algorithm first extracts the object in the image,enhances the extracted object with the traditional enhancement algorithm to obtain the multi-style oversampling data,divides the image into regions,and randomly fills the enhanced data into the regions.The multi-state oversampling algorithm can increase the richness of object regions and enhance the diversity of small objects,and the region random filling scheme can ensure the uniform distribution of small objects and enhance the space utilization of the image.(3)The A-Focus algorithm based on mean optimization is proposed to address the problem that the Focus algorithm of YOLO v5 causes too large a proportion of small object feature attributes to be lost.The algorithm uses the information of regional mean values around the pixels to obtain new feature values by mapping each regional pixel and the regional mean value.The incorporation of the mean value information can mitigate the interference of background factors on the edges of small objects and enhance the amount of small object information in the output feature layer.(4)To overcome the problem that the convolutional layer in the CSPNet structure has a small receptive field and poor feature capture ability in the feature layer,a DCSPNet structure based on the dilated convolution is proposed to build a D-YOLO network.By enhancing the receptive field area,the dilated convolution makes it easier to retain the information in the upper feature layer to the lower feature layer,thus ensuring the adequate transfer of small object features.The experimental results show that the D-YOLO algorithm has significantly improved over the YOLO v5 algorithm in small object detection,small object detection in the complex background,and dense small object detection,with a 1.65%improvement in accuracy,0.86% improvement in recall,and 1.67% improvement in mAP.更多还原