Study On Region Semantic Templates For Forensic Image Retrieval

In the public security area, surveillance cameras and visual acquisition systems are everywhere in the cities, and acquiring a lot of visual information every day for the public security department. How to get access to the information we needed efficiently from the massive public security database has become a new challenge to us. In this paper, aiming at the criminal scene investigation (CSI) image retrieval, which from the realty cases, using the content-based image retrieval system, combining with the actual situation of CSI image retrieval, and proposed an improved method for CSI image retrieval.The CSI images from the real-world cases were acquired from the Criminal Investigation Bureau of Shaanxi Province, we preprocessing and selected 400 images in 8 categories as the testing database. First, we used the content-based image retrieval system for CSI image retrieval experiments. During the experiment, the HSV color space color histogram were used as the color feature, the mean and variance of the three layer wavelet decomposition coefficients were used as the texture feature, and the above two kinds of feature fused together as the fusion feature, to experiment respectively. In the similarity measure link, the Euclidean distance and block distance effect respectively. As a control experiment, the same methods were used on the standard image database COREL for comparison. Experiment results showed that, simply use these existed image retrieval algorithms for CSI image retrieval is not effective as it is on COREL database. The analysis showed that this is result may due to the special characteristics of CSI image database that we need to pay attention to. In order to deal with influence of the specialty of the CSI image database in image retrieval process, new method needed to be developed. The experiment results also showed that, using the block distance as the similarity measure algorithm can lead to a better precision results, comparing with the Euclidean distance, which means it can evaluated the similarity between images better than the Euclidean distance both in the CSI image database and the COREL database.Secondly, based on the content-based image retrieval system, we proposed an improved system aiming at the specialty of the CSI image database:region-based semantic templates for CSI image retrieval framework. The retrieval framework contains four parts:the user submits a query image with the region of interest, build the region semantic templates, also known as the’objects’, pre-classification and image ranking. Then with these regions and the relationship of their semantic meanings we constructed a hierarch structure for the CSI image database using ontology. All these were done in order to use one or several regions or ’objects’to represent the whole image semantically, so that we can determine the class of the image by recognize the regions using the low-level feature and the relationship of the regions semantically. The experiments showed that, this method is effective in improving the mean precision of the CSI image retrieval, comparing with the content-based image retrieval system.Finally, in order to future improve the performance of the framework we added a weighting algorithm into the framework. Inspired by the idea of a weighting algorithm from the text retrieval named TF-IDF (term frequency-inverse document frequency) and information entropy, we calculated the weights of every’objects’, this weights stand for the representative and importance level to the original image it belongs. With the comparing of the weights, can be helpful in the process of pre-classification, helping to determine the class of the image it belongs to. In the calculation of TF-IDF values with information entropy can overcome some defects of the original TF-IDF, but also caused new problem that only the classes which have more than one’objects’is suited for the method because of the entropy computing. So the experiments only used the suited classes of the image for retrieval (in the whole database scale), the results showed that this method is effective generally. Although the improve extent for the mean precision is quite different among different categories.