【作者】 朱里；

【导师】 陶文兵；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2022， 博士

【摘要】 目标检测是计算机视觉领域的核心问题之一,也是解决高级视觉任务的基础。近年来,以卷积神经网络为代表的深度学习技术席卷计算机视觉领域,目标检测领域的研究在这股浪潮中取得了长足的进步。然而,在训练过程中存在的样本不平衡、样本选择不合理等问题,制约了其进一步发展,巨大的计算量和网络规模也给算法在实际场景中的落地带来挑战。为此,论文针对目标检测网络对更高检测精度和更快推理速度的需求,从训练优化和网络结构精简两个方面开展研究。论文通过探究训练过程中样本不平衡以及标签分配问题对网络的影响,优化网络训练过程,使网络在无需大幅增加网络规模的情况下实现精度提升,并针对检测网络自身特点,提出专为目标检测任务设计的通道剪枝和动态推理方法,实现检测网络的推理加速。论文主要研究内容及贡献概括如下:首先,为解决当前两阶段方法中存在的不同质量正样本数量分布不均匀的问题,提出基于交并比(Intersection-over-Union Overlap,Io U)均匀分布的目标检测算法,通过对标注框施加扰动并均匀采样,得到各Io U分段数量一致的样本,并代替原始区域生成网络的输出,来作为网络第二阶段的训练样本。该方法有效提高了高质量样本在训练过程中的占比,使得检测网络在面对高质量候选区域时能够得到定位更为精确的检测框。此外,该方法还通过更新位置回归后的候选区域的特征,解决了Io U预测分支在推理过程中遇到的特征偏移问题,增强了网络对定位质量的预测能力。其次,提出一种采用多标签分配策略的目标检测算法。通过分析当前两种主流的标签分配策略对检测网络性能以及输出特性的影响,提出一种分别利用一对多与一对一标签分配策略定义两组样本同时参与训练的方法,并设计对齐模块挖掘众多预测结果之间的关系以生成掩码,实现从一对多分类得分到一对一分类得分的转化,从而起到替代非极大值抑制算法的作用。该方法结合两种分配策略的优点使检测网络摆脱了对非极大值抑制算法的依赖,从而实现完全端到端的目标检测,并且能够有效提高检测性能。再次,对于目标检测网络的模型压缩问题,当前主要为图像分类网络设计的通道剪枝算法忽略了分类任务与检测任务的区别,无法准确定位出检测网络中每一层的关键通道,本文提出一种面向目标检测网络的定位感知通道剪枝算法,通过定位感知损失与局部重建误差来衡量每个通道对于检测性能的重要程度,帮助确定同时包含分类与定位关键信息的通道,并通过构建定位感知网络实现对主干网络和检测头网络的通道剪枝,所提出的方法在大幅压缩网络通道数的同时保持了网络性能。最后,针对目标检测网络对动态推理结构的需求,提出一种基于动态推理的目标检测算法。该算法通过构建多尺度密集连接结构实现具有动态网络深度的目标检测网络,并针对网络结构在动态调整过程中所引发的计算量和参数量急剧增加的问题,提出对多尺度特征融合和检测头部分进行结构优化,在保持性能的同时有效控制了计算量和参数量的增长。该方法可根据不同数据输入进行网络结构的自适应调整,动态减少网络结构冗余,在保证精度的同时加快网络整体的推理速度。更多还原

【Abstract】 Object detection is one of the core problems in the field of computer vision and the basis for solving high-level vision tasks.In recent years,deep learning technology represented by Convolutional Neural Networks has swept the computer vision fields,and research of object detection has made great progress in this wave.However,problems such as unbalanced samples and unreasonable sample selection in the training process restrict its further development.The huge computational load and network scale also bring challenges to the implementation of the algorithm in practical scenarios.In view of the demands of object detection network for higher detection accuracy and faster inference speed,this dissertation conducts research from two aspects: training optimization and network structure simplification.In this dissertation,the training process is optimized by exploring the influence of sample imbalance and label assignment problem on the network during training,so that the network can improve the accuracy without greatly increasing the network size.In addition,according to the characteristics of the detection network,a channel pruning method and a dynamic inference method specially designed for the detection task are proposed to accelerate the inference of the network.The main research contents and contributions of the dissertation are summarized as follows:First,an object detection algorithm based on Io U uniform distribution is proposed,which aims to solve the problem of imbalance distribution of positive samples at different quality levels in the current two-stage method.By perturbing the ground-truth and sampling uniformly,the same number of samples for each Io U segment are obtained to replace the outputs of the original region proposal network as the second-stage training samples.This method effectively increases the proportion of high-quality samples in the training process,so that the network can obtain more accurate detection boxes when faced with high-quality proposals.In addition,by updating the features of the proposal after regression,this method also solves the problem of feature offset encountered in the inference process of Io U prediction branch,and enhances the network’s ability to predict the localization quality.Second,an object detection method based on a multi-label assignment strategy is proposed.By analyzing the influence of two current mainstream label assignment rules on the detector,a method is proposed to define two sets of training samples using one-to-many and one-to-one label assignment rules for training.And an alignment module is also designed to generate masks by fully mining the relationship between the prediction results,so as to realize the transformation from one-to-many classification scores to one-to-one classification scores,thus playing the role of replacing the non-maximum suppression algorithm.The method combines the advantages of both rules.By using this,the network can get rid of the dependence on the non-maximum suppression algorithm.Thus,the complete end-to-end object detection can be realized and the performance is effectively improved.Next,for the model compression problem of the detector,the current channel pruning algorithms mainly designed for the classification network ignore the difference between the classification task and the detection task,and they cannot accurately locate the key channels of each layer in the detection network.This dissertation proposes a localization-aware channel pruning algorithm for object detection networks.The importance of each channel to the performance is measured by the location-aware loss and local reconstruction error,helping to determine the channel that contains both classification and location key information.Then,through the constructed localization-aware network,the channel pruning operation of the backbone and the detection head is realized,which greatly compresses the number of network channels while maintaining the performance of the model.At last,aiming at the requirement of object detection for dynamic inference structure,an object detection algorithm based on dynamic inference is proposed.This method enables an object detection framework with dynamic network depth by constructing a multi-scale dense connection structure,and by simplifying the structure of the multi-scale feature fusion and the detection head part,the problem of the sharp increase in the amount of computation and parameters caused by the dynamic reasoning process is solved.The algorithm can adaptively adjust the network structure according to different inputs,and speed up the overall inference speed of the network while ensuring the accuracy by dynamically reducing the redundancy of the network structure.更多还原