CSA-ASPP-Net: end-to-end laser stripe centerline extraction with joint channel-spatial attention and atrous spatial pyramid convolution

This study presents an end-to-end laser stripe centerline extraction model, CSA-ASPP-Net, which enables direct mapping from raw laser stripe images to high-precision sub-pixel centerlines through the integrated design of atrous spatial pyramid pooling (ASPP) and channel-spatial attention (CBAM). Our model addresses the limitations of traditional methods, which rely heavily on manually designed features, as well as existing deep learning approaches, which require segmentation-extraction centerlines. By innovatively integrating attention-guided feature enhancement and multi-scale contextual perception modules into an encoder-decoder architecture, the proposed model enables single-stage completion of stripe localization and refinement. The experimental results demonstrate that this end-to-end framework achieves a precision of 94.85% and an average localization error of 0.64 pixels in test images, with a processing speed of 0.15 s per image, highlighting its computational efficiency. The results demonstrate that the CBAM module effectively mitigates background interference by emphasizing salient features, while the ASPP module enhances adaptability to various stripe morphologies through its multi-scale capability. This research provides an innovative and integrated solution tailored for structured light measurement systems, combining efficiency with high precision in laser stripe processing.