融入特征交互与注意力的轻量化混凝土裂缝分割算法

Lightweight concrete crack segmentation algorithm integrating feature interaction and attention

摘要:

裂缝是混凝土建筑结构中最大的安全隐患之一。为实时高效分割混凝土裂缝并及时评估其危害,提出一种改进DeepLabV3+的轻量化裂缝分割算法。首先,使用MobileNetV3作为轻量级主干,大幅降低模型参数量；其次,使用尺度内特征交互模块(attention-based intrascale feature interaction,AIFI) 建模全局信息并引入基于归一化的注意力模块(normalization-based attention module,NAM) ,促进多层次裂缝特征信息的交互；此外,提取低层次高分辨率特征后引入混合注意力机制(ACmix),更有效捕获细节特征；最后,设计C2f-SCConv模块,对融合后的高低层级特征流解码,减小计算冗余,提升对多尺度特征的感知能力。在公共裂缝数据集Concrete 3k和Asphalt3k上的实验结果表明,提出的模型参数量相比DeepLabV3+降低了88.1%,像素准确率提升0.02%,平均交并比 (mean intersection over union,mIoU) 达到了86.21%,平均帧率为47.91帧/s,在显著降低模型复杂度的同时提高了对裂缝的分割效能。

Abstract:

Cracks pose one of the most safety hazard to concrete building structures.A lightweight crack segmentation algorithm with improved DeepLabV3+ is proposed for efficiently segmenting concrete cracks and assessing their hazards in a timely manner.Firstly,MobileNetV3 is used as the lightweight backbone to significantly reduce the number of model parameters.Secondly,the attention-based intrascale feature interaction (AIFI) module is used to model the global information,and the normalization-based attention module (NAM) is introduced to facilitate the interaction of multi-level crack feature information.In addition,the mixed model of both self-attention and convolution is introduced after extracting the low-level high-resolution features,which captures the detailed features more efficiently; and finally,the C2f-SCConv module is designed to decode the fused high- and low-level feature streams,reducing computational redundancy and improving the perception of multi-scale features.Experimental results on the public crack datasets Concrete3k and Asphalt3k show that the number of parameters of the proposed model is reduced by 88.1% compared with that of the DeepLabV3+ model,the pixel accuracy is improved by 0.02%,the mean intersection over union (mIoU) reaches 86.21%,and the average frame rate is 47.91 frames per second.It means that the proposed methods reduce complexity of the model while improve segmentation efficiency to the cracks significantly.