Research Of Shoeprint Image Fine-grained Outlier Detection Algorithm

The shoeprint is one of the significant evidences in crime scenes.For a large number of shoeprint images,constructing a dataset based on clustering algorithms for automatic management can improve the efficiency of cases detection.However,because the accuracy of clustering algorithms are low and the artificial labeling on data is subjective,the abnormal images which are different from others in a dataset will appear,and they are called outliers.Outliers will greatly affect the accuracy of post-processing on shoeprint images.The accuracy of recognition and classification algorithms could be improved by the detection and removal of the outliers in a shoeprint image dataset.Since the constructing of dataset is based on the similarity of the pattern images,there are high similarity between images in the dataset.Differences between outliers and inliers are mostly in a small independent connected area of shoeprint images,so the outlier detection of shoeprint images belongs to the fine-grained detection.Therefore,an outlier detection algorithm for shoeprint images is proposed based on analyzing the characteristics of shoeprint images.Main works are as follows:1)An outlier detection algorithm based on both local and global features is proposed.By analyzing characteristics of shoeprint images,an algorithm based on the bag of visual words is proposed to describe the information of part-regions on shoeprint images,and it calculates the outlier factors of shoeprint images by combining the information of local parts with the global features which achieves fine-grained outlier detection.The AUC of the proposed algorithm on the shoeprint images reaches to 80.20%.2)An outlier detection algorithm based on sparse self-representation and minimum spanning tree clustering is proposed.According to the fact that both outliers and outlying clusters are in the shoeprint images dataset,a detection algorithm that takes into account outliers and outlying clusters is proposed.The proposed algorithm firstly uses sparse self-representation methods to detect and remove outliers.Then,a clustering algorithm based on minimum spanning tree is used to detect outlying clusters.The experimental result shows that the proposed algorithm can detect outlying clusters while detecting isolated outliers.The AUC of the proposed algorithm on the shoeprint images reaches 83.31%.3)An adaptive outlier detection algorithm based on natural nearest neighbor is proposed.Based on the fact that the threshold is hard to be determined in the outlier detection algorithm,the natural nearest neighbor searching algorithm is used to adaptively detect outliers in the dataset according to the distribution characteristics of the dataset.At the same time,the algorithm presents a way to the problem that the inliers are mistaken as outliers because of uneven distribution of the inliers in the dataset.The experimental result shows that the proposed algorithm can detect outlying clusters without parameters compared with other clustering algorithms.The AUC of the proposed algorithm in shoeprint images dataset reaches 82.09%.