Learning-based multi-view profilometry for a telecentric structured light imaging system

Deep learning techniques have been widely applied in fringe projection profilometry. However, existing methods focus on single-view measurements. This Letter introduces, for the first time to our knowledge, an end-to-end multi-view fringe projection profilometry (MVFPP) learning framework. It treats the decoding process as a deep feature correlation, enabling adaptation to a wide range of patterns and views of input using a feature transfer algorithm, which is guided by global spatial consistency constraints. Compared with traditional multi-stage MVFPP methods, our method more stably recovers dense surface representations, even under conditions of low coding bits, low reflectivity, and shadow occlusion. To better evaluate our method, a dataset of industrial electronics scenarios has been collected. Experimental results demonstrate that the proposed method can achieve state-of-the-art results with flexible coding patterns and coding bits.