基于融合多尺度多通道特征的深度监督网络实现裂缝检测

尽管卷积神经网络浅层特征可蕴含一些细节信息,但也包含大量噪声。对于宽裂缝,浅层信息则作用不大。因此,本文提出了一个基于VGG16骨架并融合深层特征的FCN分割网络,并在每层加入侧边输出以直接监督模型。此外,我们还采用了一种名为Focal Loss的损失函数来解决数据集本身正负样本分类不平衡的问题。这种多尺度多通道深层特征与独特的损失函数融合应用,使网络具备很强的抗干扰性和更快的收敛速度。在DeepCrack数据集上,本文提出的深层特征融合网络(Deep Feature Fusion Network,DFFN)与HED、FCN和DeepCrack相比,表现出更好的性能和更快的推理速度。

Implementing crack detection using a deep feature fusion convolutional neural network

Although shallow features in convolutional neural networks can contain some detailed information, they also include a large amount of noise. For wide cracks, shallow information is not very effective. Therefore, this paper proposes an FCN segmentation network based on the VGG16 skeleton and fused with deep features, with side outputs added at each layer to directly supervise the model. In addition, we also use a loss function called Focal Loss to solve the problem of imbalanced positive and negative sample classification in the dataset. The fusion of multi-scale and multi-channel deep features with a unique loss function application makes the network highly resistant to interference and faster convergence speed. On the DeepCrack dataset, the deep feature fusion network (DFFN) proposed in this paper performs better and has faster inference speed compared to HED, FCN, and DeepCrack.