Concrete Crack Segmentation Algorithm Based on Hybrid-Attention Feature Enhancement
TL;DR
DCFYOLO, a lightweight crack instance segmentation algorithm improved from YOLO11n-seg, which enhances feature representation and multi-scale contextual fusion to improve crack detection and segmentation performance in complex scenarios is proposed.
Abstract
For surface crack detection in in-service concrete buildings, this study proposes DCFYOLO, a lightweight crack instance segmentation algorithm improved from YOLO11n-seg, which enhances feature representation and multi-scale contextual fusion to improve crack detection and segmentation performance in complex scenarios. Frequency Channel Attention (FCA) is introduced into the C3k2 units of the backbone network, where multi-scale pooling extracts multi-spectral information to enhance channel-level feature discrimination. Context Anchor Attention (CAA) is added to the PAN-FPN structure in the neck to address multi-scale feature fusion and long-range context modeling in complex scenes. A Dynamic Snake Convolution with Efficient Channel Attention (DSECA) module is constructed by combining dynamic snake convolution and efficient channel attention. Through a serial design, this module provides the multi-scale feature fusion process with both geometrically adaptive sampling and channel-discriminative optimization. Experimental results on the DeepCrack dataset show that DCFYOLO achieves 73.2% Box mAP@0.5 and 67.4% Mask mAP@0.5 with 2.47 M parameters, improving the baseline YOLO11n-seg by 2.8 and 2.7 percentage points, respectively. Ablation experiments verify the independent contribution of each improved module. The algorithm demonstrates a balance between segmentation accuracy and inference efficiency under lightweight constraints on the DeepCrack dataset, and can provide a reference for research on concrete surface crack instance segmentation.