Study And Implementation On Logo Recognization Technique Based On Use Of Image Classification And Retrieval

Logo is a significant symbol of an enterprise, and Logo recognition plays an important role for retrieving the information related to enterprises with a Logo. Especially, with the rapid development of multidedia technology, it is convenient for people to capture a Logo image with mobile phone whenever and wherever possible, so Logo recognition is becomeing a key technology for people to get the information about the enterprise that the Logo stands for. However, in fact, as the Logos are often surrounded by other objects in an image, it is hard to get satisfied results applying only image search with Logo for Logo recognition. In this thesis, a novel approach of Logo recognition is proposed, which is based on use of image classification with multi-types of features to detect the Logo region and further recognize it through image retrieval.Firstly, given an known Logo image data set with class labels, multi-type features of all images in the data set are extracted after normalizing images to the same scale, then a classifier is trained based on the extracted features. A varity of experiments have been carried out to analyze the influences of different classifiers on precison, recall and F-measure, and determine the most suitable way of setting the optimal feature combinations and dimensions on the construction of Logo calssifers.Secondly, an image with unkonwn Logo is randomly partitioned into a set of rectangular regions and each sub-region is tested using the classifier obtained previously, then the one with the maximum classification score is selected as Logo region. The experimental results demonstrate that this method is effective in detecting the Logo surrounded in the image.Finally, the technique of content based image retrieval is applied to recognize the Logo through matching the detected Logo region to Logo image with class labels in the data set. Considering the fixed structure of Logo images, a new method of similarity calculation based on spatial information is proposed in this thesis. Concretely, the Logo image is partitioned into a series of small regions, then using the model of bag of words, visual phrases with the same length are generated by computing the relative coordinate of each visual word on every sub-regions, lastly the similarity between two Logo images is calculated using the visual phrases. A range of the experiments have been done to compare the proposed method with other methods of image retrieval and discuss the influences of different parameters on the retrieval performance. The experimental results show the effectiveness of Logo image retrieval with spatial information.