Achieving high-accuracy, low-latency segmentation of infrastructure cracks on resource-constrained edge devices are a critical challenge in the operation and maintenance of intelligent transportation infrastructure. To address these issues, this study proposes a collaborative optimization framework that integrates architecture design, pruning, and quantization. First of all, a multi scale feature aggregation decoder is designed. It employs grouped large kernel convolutions, a gated attention mechanism, and an efficient upsampling module to enhance the perception of crack topology. Second, a structured pruning strategy based on Taylor expansion is proposed to remove redundant channels in a task aware manner. The model is further compressed by using hybrid linear symmetric 8-bit signed integer full quantization method. Experiments on the self-built dataset and two public datasets show that the proposed framework achieves significantly better performance than the mainstream model with only 2.4 million parameters and 4.6 giga floating-point operations per second (GFLOPs). After pruning and quantization, the parameter drops to 1.37 million, storage space is reduced by over 85%, and the mean intersection over union (mIoU) remains above 0.70. Visualization analysis and robustness tests further confirm the model's feature focusing ability and generalization performance in complex scenarios. This study provides a viable technical pathway for high precision infrastructure crack detection on resource constrained edge devices.
原文链接
分享
引用
