Enhanced YOLOv8 with Multi-Scale Dynamic Deformable Convolution and Hierarchical Dual-Attention for High-Precision Crack Detection

The segmented water-cooled crucible, serving as a pivotal apparatus within the domain of high-temperature metallurgical processing, markedly enhances thermal fatigue resistance and thermal management efficiency through its modular design in conjunction with the water-cooling system. Nonetheless, the dynamic stress concentration at the joints of the segmented structure, precipitated by mechanical assembly gaps, thermal cycling loads, and the flow of the cooling medium, can readily induce microscale long cracks. The intricate geometric structure of the segmented water-cooled crucible presents a formidable challenge to image-based crack detection methodologies. Conventional techniques grapple with the equilibrium between background interference and multi-scale crack characteristics, resulting in persistently elevated false detection rates. This paper introduces an enhanced YOLOv8 detection framework, featuring key innovations such as: MultiScale Dynamic Deformable Convolution (MDConv), which adeptly captures multi-scale features ranging from sub-pixel to millimeter-level cracks by adaptively modifying the shape and receptive field of the convolution kernel, achieving an $8.7 \%$ enhancement in micro-crack recall rate compared to the baseline model; Channel-Space Hierarchical Attention (CPMS), which filters out discriminative features at the channel level while concurrently suppressing background interference from segmented joints at the spatial level, diminishing the false detection rate to $1.5 \%$ within complex backgrounds. Experimental results demonstrate that the improved algorithm attains a detection accuracy (mAP50) of $96.7 \%$ on the dataset, an improvement of $3.6 \%$ over the original YOLOv8, and achieves real-time detection at 35 frames per second in NPU, with both false and missed detection rates falling below industry standards. This accomplishment furnishes a solution characterized by high precision and robustness for the intelligent detection of complex industrial equipment.