A multi-scale retinex implementation on FPGA for an outdoor application

Retinex is one of the well-known schemes for adaptive image enhancement under poor weather conditions. Improving the visibility of input images on a target detection system is necessary to insure detection work properly. Compared to other realization methods based on DSPs and FPGAs, this paper proposes a totally parallel structure based on FPGA to implement multi-scale Retinex for an outdoor application. Gaussian kernel is optimized with address encoding and distributed arithmetic, and parallel multi-scale convolutions are realized. No frame buffer is needed for the proposed implementation itself, so the system delay is greatly saved. Experiments show the good performance of the proposed multi-scale Retinex implementation by both subjective and objective assessments. Visibility of images suffering from fog degradation and uneven light is improved. And future work is presented in the conclusion.