Luminance value modification of degraded license plate images using support vector regression in retinex model and its application to character recognition

Abstract

There is a need for a luminance value modification method that improves the character recognition accuracy of complex deteriorated license plate images captured by security cameras with deviations in noticeable illumination light components and also adopts to human visual characteristics. In order to modify luminance value according to this need, we propose a method of applying support vector regression in the Retinex model. The proposed method estimates the illumination light component more accurately than conventional Retinex processing by using support vector regression in order to cope with the deviation of noticeable illumination light component so that it improves accuracy of luminance value modification. In order to confirm the effectiveness of the proposed method, we carried out the experiment which measures the accuracy of recognizing characters by normalized cross-correlation. As a result, the correct rate of the proposed method has increased by about 39% at maximum in comparison with the conventional Retinex processing. Moreover, we confirmed that the image of the proposed method has a lower Halo than the image of the conventional Retinex processing, so that it is found that the proposed method can restore visually natural images from the viewpoint of Halo at least. Furthermore, we confirmed that the proposed method has superiority over the conventional Retinex processing even when the security camera image in the real environment is targeted. It is also found that the proposed method is more effective than the conventional Retinex processing in the application of luminance value modification methods to character recognition by convolution neural network.