Research On Application Of Relevant Recognition Technology For Supermarket Labels Photographed By Mobile Phone

Through the analysis of the characteristics of label images, this article is mainlyconcerned about the preprocessing and corresponding recognition technologies for thelabel images and the characters in it. In the part of image preprocessing, firstly thisarticle analyses the characteristics of label images, presents a series of preprocessingsteps, and does research on technologies that applied in every step. In the part ofcharacter recognition, after comparing and analyzing some common recognitiontechnologies, this paper introduces the BP neural network pattern recognitiontechnology, then analyses and designs it in details. This article carefully analyses thekey and difficult problems in the two parts, and the specific work includes thefollowing two parts：1. In the preprocessing module, it includes three parts: In the part of decreasingthe effect of non-uniform illumination, after comparing the different processingeffects between the spatial filting and frequency domain filting, this article uses thehomomorphic filtering to process the images which are affected by non-uniformillumination. In the part of binarization, firstly this article analyses and compares thecommon methods in the global and local threshold binarization. For the label imagewith colourful background, the Otsu method combined with gray stretch and the graygradient method to enhance the effects of binarization is presented. In the part ofimage denoising, after comparing and analyzing typical denoising methods, accordingto the characteristics of label images, this article uses morphological filtering todenoise label images and optimizes the method of selecting structural element. It hasobtained certain effect by experimental contrast in the above three aspects, and it laysa certain foundation for the following character feature extraction.2. In the character recognition module, firstly, this article summaries the commoncharacter features and character recognition methods. In the part of character featuresextraction, on the basis of normalizing the characters in the images, this articlepresents using an combined feature which is composed of the grid feature andprojection feature to be the combined feature vector. In the part of designing the BPneural network, this article studies on the key and difficult problems in the course ofdesigning the network in details, it includes network framework design, parametersdesign and so on. Due to the problem that the gradient descent algorithm is easy to fallinto local minimum, this article introduces the momentum-adaptive learning rate method to try to solve this problem. Finally, this paper verifies the feasibility andeffectiveness of the whole design of BP neural network through the experiments.