多尺度注意力与特征融合的深度监督路面裂缝分割算法

Deep supervised pavement crack segmentation algorithm based on multi-scale attention and feature fusion

摘要:

针对路面裂缝检测在复杂背景、裂缝像素比例不平衡情况下会出现裂缝分割不完整、漏检的问题，本文提出一种多尺度注意力与特征融合的深度监督路面裂缝分割算法。首先，利用ECA(efficient channel attention)注意力机制对EfficientNet-B3进行改进并将其作为网络模型的编码部分，以加速对裂缝像素的关注；其次，设计了多尺度通道注意力模块(multi scale channel attention module，MSCA)，其运用空洞卷积的级联并行策略提取关键性的上下文信息并增强对细小裂缝的感知能力；最后，在辅助网络部分以特征金字塔方式集成多个侧面特征图，引入深度监督机制，加快了模型收敛和提升裂缝检测的效果。在CRACK500、CFD、DeepCrack数据集上进行实验，检测结果中F1分别可达到75.87%、66.80%、86.46%，与目前先进的裂缝分割方法相比表现更佳，具有一定应用价值。

Abstract:

Aiming at the problem of incomplete crack segmentation and missed detection in pavement crack detection under complex background and unbalanced crack pixel ratio, this paper proposes a deep supervised pavement crack segmentation algorithm with multi-scale attention and feature fusion. Firstly, the EfficientNet-B3 is improved by using the ECA (efficient channel attention) attention mechanism and used as the encoding part of the network model to accelerate the attention to crack pixels. Secondly, a multi-scale channel attention module (MSCA) is designed, which uses the cascade parallel strategy of dilated convolution to extract key contextual information and enhance the perception of small cracks. Finally, multiple side feature maps are integrated in the auxiliary network in a feature pyramid manner, and a deep supervision mechanism is introduced to accelerate the convergence of the model and improve the effect of crack detection. Experiments are carried out on the CRACK500, CFD, and DeepCrack datasets. The F1 in the detection results can reach 75.87%, 66.80%, and 86.46% respectively, which is better than the current advanced crack segmentation methods and has certain application value.