Research On Intrinsic Image Decomposition Algorithm And Its Applications

Intrinsic image decomposition is a fundamental problem in computer vision, given an input image, corresponding reflectance and lightness image need to be decomposed. Early algorithm Retinex simulated the eyes’ invariance to illumination, and achieved good results, but it didn’t use the eyes’ ability to identity depth. In our work we proposal to combine the depth information with derivative Retinex, it avoids the loss of depth edge information in high level pyramids used by Stereo Retinex, which combines depth information with multi-level Retinex. Our method classifies derivatives to two kinds, one is reflectance derivatives the other is lightness derivatives, once the derivatives are correctly classified, the reflectance and lightness image can be calculated using image integration. The experimental results show that this method improves the precision of original Retinex algorithm.We also combined Retinex with SIFT algorithm and proposed a new image feature extraction and match method R-SIFT (Retinex SIFT), image feature extraction and match methods like SIFT fail when two images are taken under very different illumination, but experimental results show that R-SIFT performs well in these conditions, which means R-SIFT has better invariance to illumination.In the final part of our work we give intrinsic image decomposition algorithm’s application in license plate recognition, license plate recognition is an important part in intelligent traffic systems, in which license plate location is the basis of car plate recognition, its performance greatly influences the whole LPR system. Original license plate location methods fail in bad illumination conditions like the night condition. In our paper we combine Retinex with original license plate location method. We use the illumination invariant reflectance image to perform license plate location methods. And it solved the failure of original methods like the projection analysis method in bad illumination conditions. Based on the license plate location method we use R-SIFT match method to identify suspicious cars and gives an early warning, first we use the registered car photos as standard to locate the car plate and performs R-SIFT feature extraction method on it, then we dispatch the standard features to different monitoring points, which performs the feature extraction procedure on cars detected, and finally it matches the detected feature with the standard feature it received, when the match number reaches a certain threshold, the system gives a warning that the suspicious car is detected, and different thresholds can be used to represent the importance of the event.