The Illumination Compensation Algorithms In Pattern Recognition

With the development of computerscience, the era of Big Data and the boom of deep learning, object recognition technology has made a breakthrough. Before that, the performance of object recognition, especially face recognition system is restricted by the light condition. Many scientists also have lots of research in this field, but the results are still not ideal. However,it’s the significant step to solve the problem of object recognition in complex illumination conditions. In this context, firstly, this paper based on Retinex model, has presented a improved algorithm, AuReH(Auto color enhancement algorithm based on Retinex model in HSI color space). Then, combine Convolutional Neural Network and AuReH algorithm, present the RetiNet model, which is a end-to-end model. And the results show that this model is robust in object recognition in complex illumination conditions.This paper put emphases on researching robust object cognition algorithm based on convolutional neural network in complex illumination conditions, and improve the traditional multiscale Retinex Model. The main work is as follows:1. There many traditional color enhancement algorithms, including the spatial domain image algorithm and frequent domain image algorithm. Since Edwin Land et al[1] presented the Retienx model, different forms of Retinex algorithms have emerged, which has made significant contributions to image enhancement. In this paper, we present an improved color enhancement algorithm AuReH(Auto color enhancement algorithm based on Retinex model in HSI color space). Compared with these traditional Retinex algorithms, it has a good performance in color enhancement and image defogging.2. To solve the problem of object cognition in complex illumination conditions, we combine Convolutional Neural Network and AuReH algorithm, and present the RetiNet model, which is a illumination robust and end-to-end model.3. Design and implementation of a cross platform convolutional neural network system.4. Using the RetiNet model to train the different data sets, and analyze the results.