Research On Scene Tire Tread Pattern Retrieval Algorithm

The scene tire tread pattern retrieval is referred to output the ranked list of database patterns sorted by using the useful information of tire tread pattern images under certain discriminant conditions. In recent years, hierarchical feature extraction and unsupervised learning process of the human brain have been studied as a popular topic. The process of rapidly recognizing objects despite substantial appearance variation can be viewed as an unsupervised learning process of the nerve cells from the retina to the temporal lobe cortex area in the brain, which is also a process of hierarchical feature extraction from simple and concrete levels to complex and abstract ones. Therefore, the hierarchical feature extraction and unsupervised learning of the human brain are applied to scene tire tread pattern retrieval. A quick and effective scene tire tread pattern retrieval algorithm is proposed, which is based on sparse representation and probabilistic latent semantic analysis. Main contributions of the thesis are as follows.1) A preprocessing algorithm of tire tread pattern images is proposed.Firstly, the orientation of tire tread pattern images is normalized and the contrast is adjusted using histogram equalization method. Secondly, tire tread pattern images are filtered out the noise using nonlocal average filter. Finally, tire tread pattern images are divided into blocks, and each image block is binarized using OTSU. The experiments show that the tire tread pattern images that are preprocessed can be well separated from backgrounds, and can effectively represent the pattern information of images.2) A novel Gabor basis based primary visual feature extraction algorithm is proposed.Firstly, the Gabor basis vocabulary is constructed using Gabor kernel functions with different scales and orientations according to features of tire tread pattern images. Secondly, coefficient vectors of Gabor bases are solved through matching pursuit. Finally, the above coefficient vectors are weighted through the human-computer interaction. The primary visual features of tire tread pattern images are the weighted coefficient vectors. The experiments show that primary visual features are robust to the interferences such as partial, noises and distortions, etc. And it also correlates well with human’s subjective opinions.3) A probabilistic latent semantic analysis based topic feature extraction algorithm is proposed.Firstly, primary visual features of tire tread pattern training images are mapped to the probabilistic latent semantic analysis model to extract topic features of tire tread pattern training images and latent semantic space. Secondly, a topic dictionary is constructed from training tire tread pattern images. Finally, topic features of query images are extracted based on the latent semantic space of training images. The experiments show that topic features can well represent common characteristics of similar pattern and differences of non-similar pattern, and reduce the semantic gaps between the primary visual features and high-level semantics.4) A novel sparse representation based retrieval strategy is proposed.Firstly, the query images are sparsely coded using the constructed topic dictionary on training images. Secondly, the sparse code residuals and coefficients are collaboratively used to compute the similarity measure between the query image and a database pattern, which can reduce the effect of feature errors. Finally, the database patterns are sorted from the most similar one to least similar one. The experiments show that the proposed retrieval strategy can effectively eliminate the influence of feature errors.The comprehensive experiments show that the precision ratio of the proposed algorithm within the first 0.2% of the retrieval results on synthetic test images is more than 99.93%; the average recall ratio within the first 2.42% of the retrieval results on the real scene images is 100%, and the average precision ratio is more than 71.40%; and the proposed algorithm has better comprehensive performance than the state of the art algorithms on the recall ratio, the precision ratio and computation cost. The proposed algorithm composed of three levels is robust to interferences such as incompleteness, occlusions and noises, and has better performance on the real scene tire tread pattern images. The proposed algorithm has been successfully used in practical applications.