【作者】 胡涛；

【导师】 许贵平；

【作者基本信息】 华中科技大学 ， 计算机技术， 2020， 硕士

【摘要】 最近几年,目标检测算法模型层出不穷,取得了令人瞩目的进展,在工业领域和生活领域都找到了许许多多的适用场景。但现有的目标检测模型,大多数都是针对通用的自然场景下的图像目标来设计解决方案。对于航拍图像中的小目标来说,检测效果不是很理想,存在着巨大的挑战。因此,论文着手于航拍图像的实际场景,从浅层特征与深层特征融合,感受野增强,轻量级特征提取网络等角度来解决航拍图像小目标定位与分类的问题。主要工作如下:为了实现航拍图像中小目标物体的精准检测,在单点多框检测模型SSD(Single Shot Multi Box Detector)的基础上,提出了一种基于多尺度特征融合和感受野增强的航拍图像目标检测模型MF-SSD,该模型构建了一个浅层特征两级融合,中间层特征跳跃连接融合的特征融合子网络,针对航拍目标尺度较小,外观信息较少导致的目标检测精度低问题,MF-SSD结合多个尺寸大小的特征图信息,有效地提升了小目标检测的精度。同时,在特征融合子网络的基础上,提出了一种基于空洞卷积的多分支Inception结构D-Inception来作为感受野增强模块,增强了预测特征层的局部信息表达能力,能够更好地提取目标特征和适应目标的尺度变化。为了解决特征融合和感受野增强处理后带来的模型计算复杂度过高的问题,改用轻量级的卷积神经网络来替换基础网络提取目标物体特征,可以在降低网络模型复杂度的同时,仍然保持较高的检测精度。在航拍数据集NWPU VHR-10上进行测试,实验结果表明,改进后的目标检测模型MF-SSD取得了优于原始SSD模型和Faster R-CNN模型的效果,能够高效地对航拍小目标物体进行检测,也能很好的泛化到通用目标检测中。更多还原

【Abstract】 In recent years,the target detection algorithm models emerge one after another,and remarkable progress has been made in target detection.Many applicable scenarios have been found in both the industrial field and the life field.However,most of the existing target detection models are designed for image targets in common natural scenes.For the small targets in aerial images,the detection effect is not very ideal,and there are great challenges.Therefore,this thesis starts from the actual scene of aerial images,and solves the problem of positioning and classification of small targets in aerial images from the perspectives of the fusion of shallow features and deep features,the enhancement of sensory field,and the lightweight feature extraction network.The main work is as follows:In order to realize the accurate detection of small target objects in aerial images,on the basis of SSD(Single Shot Multi Box Detector)detection model,an aerial image target detection model which is named MF-SSD based on multi-scale feature fusion and sensory field enhancement is proposed.This model constructs a feature fusion sub-network,it designs a two-level fusion strategy based on shallow features and a jumping connection fusion strategy based on middle features.The accuracy of target detection is low due to the small size of aerial target and less appearance information.MF-SSD combines with the feature graph information of multiple sizes,so the accuracy of small target detection is improved effectively.Meanwhile,on the basis of the feature fusion sub-network,a multibranch Inception structure which is named D-Inception based on dilated convolution is proposed as a sensor field enhancement module,which enhances the local information expression ability of the prediction feature layer and can better extract the target features and adapt to the target scale changes.In order to solve the problem of high computational complexity of the model caused by feature fusion and sensory field enhancement,a lightweight convolutional neural network is used to replace the basic network to extract the features of the target object,which can reduce the complexity of the network model while still maintaining high detection accuracy.The test was carried out on the aerial photography data set NWPU VHR-10,and the experimental results show that the improved target detection model MF-SSD has achieved better results than the original SSD model and Faster R-CNN model.It can detect the small target object in aerial photography efficiently and can be well generalized to the general target detection.更多还原