Hyperspectral imaging (HI) technology allows us to obtain the spatial and spectral information of target objects, thus having important and various application prospects among remote sensing detection, military reconnaissance, and other fields. Recently, snapshot HI systems have been favored widely due to their advantages including the miniaturization and the lightweight, to maintain the mainstream of HI technology research. D. S. Jeon et. al. used a single diffractive optical element (DOE) instead of multiple complex elements to modulate the diffractive wavefront which greatly reduce the system volume and be significant application potential in portable equipment and small payloads1. However, the minimum line width of surface micro structure on the DOE is restricted by the current processing capability, which has a negative impact on the wavefront modulation. Consequently, the high-performance DOE design method considering for machinability is an urgent problem to be solved. We proposed a high-performance processable DOE design method which utilizes the physical model to constrain the minimum line width in the DOE surface. By using the end-to-end deep learning method, achieving processable design of DOE has almost no impact on its performance and imaging quality of hyperspectral imaging systems. Proposed model overcomes the processing problem of DOE from a design perspective without sacrificing their functionality. Comparing to the results without the constraint of the physical model, the proposed physical model has significant effects in constraining surfaces.
|