Address Block Localization For Chinese Postal Envelopes With Clutter Background

The address block localization also known as automatic address block localization, is an integral part and the key technologies of automated postal sorting machines. Usually, the background of plastic or advertising envelopes are clutter and complicated, thus the research of address block localization about them has great application value.Many methods have been proposed for AABL. But most of them are based on knowledge, i.e. according to the statistical characteristics of the envelopes with binarization to identify the destination address block. Such methods are simple, strictly dependent on the prior knowledge and statistical features, thus the results are not satisfactory. The destination address block is regarded as a whole pattern with the method based on structure, through a large number of sample to model the features, according to the similarity measure to determine whether the destination address block is. In this paper we proposes two structure-based localization methods.The first envelope destination address localization method is based on Histogram of Oriented Gradient referred to as HOG and Conditional random fields referred to as CRF. As a non-candidate based localization method, we obtains a series of windows with the same size and overlapping each other by the way of sliding window, and then transfer the localization task to binary classification problem. The gradient feature is extracted by HOG descriptors. The spatial structure information and gradient feature are integrated by CRF through joint learning CRF parameters and dictionary. We calculate the probability of block-level map, which is the probability of each candidate block belonging to the destination address block, and then get the pixel level probability by interpolation.The second method is an address block localization method based on vision saliency, ABLVS for short. Firstly, we use the binarized normed gradients method to identify the candidate regions, whose local saliency maps are similar with the training blocks’; Secondly, region covariance descriptors are used to nonlinearly fuse various low-level features, for example, pixel location, intensity, gradient and texture features; Lastly, we use support vector machine to classify the regions into address or non-address block. In order to further pinpoint the address block’s text region in envelopes, we use the image signature to compute sparsely saliency, then apply Gaussian filter to smooth the saliency map. Compared to the first method which is suitable for all types of envelopes, this method is more suitable for envelopes with clutter background, i.e. advertising envelopes.Experimental results show that both of them can accurately detect the destination address block and has strong robustness to the changes in illumination, rotation, and clutter background.