A car-face region-based image retrieval method with attention of SIFT features

Abstract

A traffic still image captured by the high-definition camera located in traffic block port often contains several vehicles. It provides an important clue to solve vehicle crime cases to retrieve all these images that contain the given type of car. To enhance the performance of image retrieval, we proposed a car-face region-based image retrieval method with the attention of SIFT features. In our method, the first step is to find all the car-face regions from an original traffic image, and this original image is represented as a set of car-face regions. Secondly, a similarity measure metrics is proposed with a light intense training of the attention value of SIFT key points on a very small identified images set. Finally, according to the similarity between the input region and the target region, all the images with at least one similar region are retrieved. We carry out this method on a training set of 100 positive car-face region-images. Compared with the famous training-based SVM method, our method achieved higher precision at the same recall with lower training intensity.