In the realm of industrial vision, depth estimation is a cornerstone technology, yet existing approaches continue to grapple with significant challenges. While this technology enables precise quality control, robotic navigation, and three-dimensional reconstruction, issues persist with reflective surfaces in industrial settings and occlusion processing, impeding widespread adoption. We begin by highlighting the current limitations of depth cameras in generating depth maps, including high noise levels, indistinct edge textures, and the degradation of depth detection caused by specular reflective objects. To solve these problems, we propose the WaveSamba method. At its core is a wavelet-based Mamba architecture (WaveSamba) decoding head, which ingeniously leverages foundation models to integrate prior knowledge from CLIP and SAM, resulting in a metric depth estimation detection network. WaveSamba achieves fine-grained, instance-level depth estimation and effectively mitigates the noise and reflection problems inherent in industrial depth cameras. This marks the first attempt to in-corporate pretrained model prior knowledge into depth estimation for industrial applications. Our experimental findings demonstrate that our approach surpasses existing camera algorithms in producing visible depth maps, addressing the pervasive issues of imaging noise, edge blurring, and reflections in depth cameras. This study advances the application of depth estimation in industrial settings and offers fresh perspectives on combining diverse visual representations to tackle complex visual tasks.