毕业生知识库

      目标检测作为计算机视觉领域的核心任务之一，在现实生活中拥有广泛的应用场景。目前，目标检测模型越来越倾向于采用更深层、复杂的网络结构，以换取更精确、稳定的检测效果。然而，随着移动互联网和移动设备的普及，越来越多的应用场景要求目标检测模型能够在边缘计算平台上部署，并且保持符合应用需求的检测精度。边缘计算平台的主要特点是计算资源有限且应用场景实时性强。因此，如何设计一个轻量、高效的网络模型以满足边缘计算平台的需求成为一个亟待解决的关键问题，这也促使了轻量级目标检测模型的研究。

      近年来，面向目标检测模型的轻量化算法取得显著的研究进展,如轻量化模型架构、网络剪枝、知识蒸馏、参数量化、低秩共享等,这些方法为视觉轻量化模型的理论和应用提供了重要支持。网络剪枝和知识蒸馏可以显著减少模型规模和计算复杂度，从而适应资源受限的应用场景，但仍然面临一些需要解决的问题。网络剪枝在剪枝的粒度和准确性上很难平衡，细粒度的剪枝可以很好地保持精度，但需要边缘平台的硬件加速；粗粒度的剪枝方便实施和普及，但很难保持精度。传统的知识蒸馏主要适用于分类模型，使用大模型的分类概率作为软标签训练小模型，但目标检测模型的预测结果为检测框中心及宽高信息。如果直接将其应用于目标检测模型蒸馏，无法起到软标签的作用。因此，本文针对面向边缘计算平台的轻量级目标检测模型，对其网络剪枝难以平衡粒度及准确性和知识蒸馏难以传递目标信息两个方面进行了研究。

（1）基于滤波器动态筛选的重整剪枝

      提出一种基于滤波器动态筛选的重整剪枝方法。该方法的关键思想是使用三个策略实现高精度的滤波器剪枝,即滤波器动态筛选、转移滤波器信息、参数重整剪枝。首先，通过权重欧式距离聚类提取结构间依赖关系，结合权重大小表征的重要性做滤波器动态筛选。其次，借助辅助卷积层实现被剪枝滤波器信息转移。最后，使用参数重整实现间接的滤波器剪枝。实验结果表明，通过应用所提综合剪枝策略，可有效保留模型功能多样性，并进一步提高滤波器剪枝的效果。与现有基于参数重要性滤波器的剪枝方法相比，基于滤波器动态筛选的重整剪枝方法可充分挖掘滤波器间潜在信息，能更加全面地评估滤波器的重要性，并可有效保留滤波器功能的多样性。此外，该方法有效地转移了被剪枝结构信息，并通过参数重整间接地实现了滤波器剪枝，可更好地保留模型的有效信息和检测精度，为网络模型剪枝提供了平衡剪枝粒度与模型检测精度的新方案。

（2）基于分布范围损失的知识蒸馏

      提出一种基于分布范围损失的知识蒸馏。该方法将检测框损失分解为中心位置分布损失、目标框面积分布损失和密度分布损失三个部分，进一步结合目标类别损失以及目标存在性损失，构成分布范围损失。该方法的关键思想是将检测框的硬标签转换为语义上的软标签，即所传递的信息为检测框的分布范围损失而非具体的检测框损失。实验结果表明，采用本文提出的目标检测模型蒸馏方法，能够有效提高模型在目标检测任务中的性能。与传统的硬标签损失相比，分布范围损失更加注重模型对目标位置、尺寸和密度的理解和学习，使得模型更具鲁棒性和泛化能力。此外，该方法还能有效地减轻模型过拟合的问题，提高模型在复杂场景下的泛化性能，进而推动目标检测技术在实际应用中的发展和应用。

    As one of the fundamental tasks in the field of computer vision, object detection has a wide range of application scenarios in real life. Currently, there is an increasing trend towards adopting deeper and more complex network structures for object detection models to achieve higher accuracy and stability. However, with the popularity of mobile Internet and mobile devices, there is a growing demand for object detection models that can be deployed on edge computing platforms while maintaining sufficient accuracy to meet application requirements. The main characteristics of edge computing platforms are limited computing resources and strong real-time application scenarios. Therefore, how to design a lightweight and efficient network model to meet the needs of edge computing platform has become a key problem to be solved, which also promotes the research of lightweight object detection model.

    In recent years, significant research progress has been made in lightweight algorithms for object detection models, such as lightweight model architecture, network pruning, knowledge distillation, parameter quantization, low-rank sharing, etc. These approaches provide important support for the theoretical interpretation and application of visual lightweight models. Knowledge distillation and network pruning can significantly reduce the scale and computational complexity of the model, so as to adapt to resource-constrained application scenarios, but still face some problems that need to be solved. Network pruning is difficult to balance the grain size and accuracy of pruning. Fine-grained pruning can maintain accuracy well, but it needs hardware acceleration of edge platform. Coarse-grained pruning is convenient to implement and popularize, but it is difficult to maintain accuracy. Traditional knowledge distillation is mainly applicable to classification models, using the classification probability of a large model as a soft label to train a small model, the prediction result of the object detection model is the center of the detection frame and the width and height information. If it is directly applied to the distillation of the object detection model, it cannot play the role of soft label.Therefore, in this thesis, we study the difficulty of knowledge distillation to transfer target information and network pruning to balance granularity and accuracy of lightweight object detection models for edge computing platforms.

(1) Renormalization Pruning Based on Dynamic Filter Selection

    A renormalization pruning based on dynamic filter selection is proposed . The key idea of this method is to achieve high precision filter pruning by using three strategies: dynamic filter screening, transfer filter information and pruning with parameter renormalization.Firstly，the dependencies between structures are extracted by clustering the euclidean distance of the weights, and the filters are further selected by combining the importance of numerical representation of weights.Secondly，the information transfer of the pruned filter is realized by the auxiliary convolution layer. Finally，indirect filter pruning is achieved using parametric renormalization.The experimental results show that by applying the proposed comprehensive pruning strategy, the functional diversity of the model can be effectively preserved and further improve the effect of filter pruning.Compared with the existing filter pruning methods based on parameter importance, renormalization pruning based on dynamic filter selection can fully explore the potential information among filters, evaluate their importance more comprehensively, and effectively preserving the diversity of filter functions. In addition, the proposed method can effectively transfer the pruned structure information, and indirectly realize the filter pruning through parameter renormalization, so as to better retain the effective information and detection accuracy of the model, and provide a new scheme for the balance of pruning granularity and model detection accuracy.

(2) Knowledge Distillation Based on Distribution Range Loss

    A knowledge distillation based on distribution range loss is proposed.This method divides the detection frame loss into three parts: center position distribution loss, target frame area distribution loss and density distribution loss, and further combines the target category loss and target existence loss to form the distillation loss based on distribution range. The key idea of this method is to convert the hard label of the detection frame into a soft label in meaning. That is, the information transmitted here is the loss of the distribution area of the detection frame, rather than the specific detection frame loss. The experimental results show that the distillation loss of object detection model proposed in this thesis can effectively improve the performance of the model in object detection tasks. Compared with the traditional hard label loss, distribution range loss pays more attention to the understanding and learning of the target position, size and density, and makes the model have better robustness and generalization ability. In addition, the method can effectively reduce the overfitting problem of the model, improve the generalization performance of the model in complex scenarios, and thus help promote the development and application of object detection technology in practical applications.