An improved attention mechanism is proposed in this content, which captures both channel attention and spatial attention through a three-branch structure. The approach introduces cross-dimension interaction to explore the interplay between spatial dimensions and the channel dimension of the input tensor. This method enhances crack detection by aggregating features in both horizontal and vertical directions. The content also discusses the Mixed Strip Convolution Module (MSCM), which captures long-range dependencies in crack information from four directions. Additionally, the Receptive Field Enhance (RFE) module is introduced, featuring four branches with the same hole rate of dilated convolutions to expand the receptive field. A weighted binary cross-entropy loss function is proposed to address the challenge of imbalanced crack and non-crack pixels in the dataset during training. This loss function assigns different weights to crack and non-crack pixels to prevent the model from focusing solely on the majority class. Overall, these modules and techniques aim to improve crack detection performance by enhancing feature aggregation, capturing long-range dependencies, expanding the receptive field, and addressing class imbalance during training.
Source link
Source link: https://www.nature.com/articles/s41598-024-65909-1
#Improving tunnel crack detection with linear seam and attention mechanism
