Partial Occluded Logo Recognition Based On Sparse Coding

Logo can be seen everywhere in our daily life. It provides enough information to identify the brands of products. And logo recognition technology is widely used in the detection of brand on electronic business platforms and the illegal use of trademarks. Therefore logo recognition technology has great value in practical application. However, to apply logo recognition in natural scenes, there are still many problems to overcome, such as the dramatic scaling, occlusion problem, lighting changing, and affine deformation. Among these problems, the occlusion problem is a challenging issues in a real scenario, for the reason that logos often do not fall entirely within the camera lens and they are easy to be occluded by other objects. Currently, partial logo recognition has not attracted much interest. This paper presents a partial logo recognition framework, which is capable of handling logo occlusion. The main work and contributions are as follows:(1) A complete and effective partial logo recognition framework is proposed. Existing logo recognition algorithms such as Bo W(bag of words model), are under the assumption that the logo should not be significantly occluded. If the assumption does not hold, the performance of the algorithm will be significantly reduced. This paper propose a new framework for solving partial logo recognition problem. In this framework, we use Faster R-CNN to detect partial logo, and apply SRC(Sparse Representation Classification) algorithm to partial logo recognition.(2) A new dataset for partial logo recognition is proposed as a new benchmark. The method can be used to evaluate the performance of the algorithm under different degrees of occlusion. Results on the generated occluded image data sets show that our method outperforms the Bo W model with an average accuracy 92.2% and exceeds 6.5%. Under different occlusion ratios, sparse coding method used in this paper performs better than Bo W algorithm. The algorithm is proved to be robust to occlusion.(3) A new method of dividing overlapped sub-region is proposed. When dealing with the occlusion problem, the common skill is dividing the image into equal parts and then fusing the classification results of the sub-region. But this approach does not take the different weights of the sub-regions into account. In fact, the possibility of a central region containing logo information is higher than other regions. This paper divides a logo image into overlapped sub-regions to avoid the situations in which there are too less key points. And this method assigns more weights to the center region. Experiments show that the method is better than the equal region weights method.(4) The color feature is adaptively used in this framework. Logos usually have eye-catching color and shape, existing logo recognition methods often use the more stable texture feature, and less use color information. Although the color of some categories of logos varies greatly in different scenes, there are still some kinds of logos which has stable color distribution. When the logo texture is too simple, and the texture features classification results are unreliable, we do not use SIFT(Scale-Invariant Feature Transform) feature to classify logos and concatenate HSV(Hue, Saturation, Lightness) color histogram with SIFT feature as a new feature instead. The experiment results show that this method improves the recognition results compare to the method which utilities local feature only. It can serve as a supplement to texture feature.