摘要： 由于没有跳跃连接结构的深度神经网络在超过一定深度后难以训练，近期的深度神经网络模型大都采用跳跃连接结构来解决优化问题和提高泛化性能。然而，人们对于跳跃连接结构如何影响深度神经网络特征提取的理解还较少，大多数情况下，这些模型仍然被认为是黑盒。为了分析跳跃连接结构对深度神经网络特征提取的影响，本文从特征可视化的角度出发，以基于扰动的方法为切入点，提出了一种在保持图像总体颜色分布和轮廓特征基本不变的前提下弱化图像细节特征的扰动方法，并将其命名为网格乱序模糊方法。同时，研究结合特征可视化中的激活最大化方法和所提出的网格乱序模糊扰动方法，分析了拥有不同程度跳跃连接结构的经典图像分类深度神经网络模型VGG19, ResNet50和DenseNet201，结果表明：第一，没有跳跃连接结构的深度神经网络只提取了图像中较强的特征，提取的特征数量比较少，而拥有跳跃连接结构的深度神经网络提取了图像中更多的特征，但是这些特征相对较弱；第二，跳跃连接结构让模型更关注图像的局部颜色分布和全局总体轮廓，而不过多依赖图像细节特征，并且跳跃连接结构越密集，这种趋势越强。

Abstract: Due to the difficulty in training deep neural networks without skip connection structures when they exceed a certain depth, skip connection structures have been integrated into most of the recent deep neural network models to address optimization issues and enhance generalization performance. However, our understanding of how skip connection structures affect the feature extraction in deep neural networks is still limited, and in most cases, these models are still considered as “black boxes”. In order to analyze the impact of skip connection structures on feature extraction in deep neural networks, this paper focuses on perturbation-based methods and introduces a method named Grid-shuffled Blurring, whose aim is to reduce the fine-grained details within an image while maintaining its overall color distribution and contour characteristics as much as possible. Meanwhile, this paper employs the Activation Maximization method in feature visualization and the Grid-shuffled Blurring perturbation method to analyze classic deep neural network models like VGG19, ResNet50 and DenseNet201 in image classification tasks, which are with different levels of skip connection structures. The experimental results show that firstly, neural networks without skip connection structures only extract stronger features from images, resulting in fewer extracted features, while neural networks with skip connection structures extract more features from images, albeit relatively weaker ones. Secondly, skip connection structures make the models focus more on local color distribution and global contours of images, rather than relying too much on detailed features on images, and the more skip connection structures there are, the stronger this trend becomes.