Research And Implementation Of Image Enhancement Algorithm Based On Night Images

Recently, with the increasing popularity of mobile products such as smart phones and the tablets, the ordinary life can be recorded at any time, also interesting events could be shared with each other more conveniently. However, there exists a dilemma in taking pictures at night due to the hardware limitation of devices and the illumination restrictions of shooting scenes. In other words, the picture taken with flash appears clearly at the cost of the destruction of ambient illumination. While the picture taken without flash appears noisy and blurred. Therefore, this thesis proposes an image enhancement algorithm based on night images.Firstly, the thesis proposes a night image registration algorithm based on Affine-SIFT: 1) the images taken with flash and without flash are preprocessed separately with histogram equalization manipulation; 2) the keypoints are detected and matched with Affine-SIFT, and the incorrect matched keypoints are removed by Random Sample Consensus algorithm. It should be determined whether there is the problem of overfitting according to the remaining matched keypoints. If there exists the problem, the adaptive matched keypoints equalization algorithm should be applied to make the location of matched keypoints more uniformly; 3) compute the transformation matrix and finally match the images with the matrix.Next, for the image fusion module, the thesis puts forward an image fusion algorithm based on detail reconstruction: 1) the algorithm converts the images into YUV color space and extracts detail layers from the flash and no-flash images in the YUV channels separately. These detail layers are used to rebuild the final detail matrix thus to obtain better detail; 2) the no-flash image is denoised via edge-preserving filter(WLS filter); 3) transfers the final detail matrix to the no-flash filtered image so that the final image enhancement result is achieved.At last, we compare our proposal with the image registration based on Affine-SIFT algorithm, the experimental results show that the proposed image registration algorithm can not only detect more keypoints, but also solve the overfitting problem caused by keypoints which are too concentrated, So the accuracy of image registration is improved.Furthermore, the proposed image fusion algorithm can maintain ambient illumination in the no-flash image, improve the details and make the image more clearly simultaneously. In addition, our proposal has such an excellent universality that it can not only reduce color distortion effectively, but also adaptively extract good details according to different environments. The proposed algorithm can be widely used to improve the quality of image which is taken at night by mobile products.