The Research Of Barcode Localization Based On Regional Gradient Statistical Analysis And Convolutional Neural Network

After many years of development, barcode technology has been greatly applied in some fields as logistic, retail business, supply chain management because of its cost-saving property and easy for application. With the popularization of smartphone and mobile internet, the usage of barcode technology is converting from simple cargo labels to much larger applications.In industrial field, accurate and robust barcode localization technology has been greatly demanded and has to be answer to the requirement of speed in mass-manufacture. Traditional image processing methods have been applied into this problem to get the location of the barcode in given images. Under certain circumstances like carefully designed lighting environment and accurate focusing, they can perform well, but often failed to handle the images captured with scale changed or defocused. What’s more, recent researches pointed out that the limit of artificial features can be easily overcome by machine learning or deep learning technique, which is able to get the internal statistical regularities of patterns from large scale data.Based on the survey of existing technologies and theories, we had an in-depth study towards the image processing and machine learning methods, and found ways to solve barcode localization problem. Our main research contributions are as follow:1. We had an intensive study on the published works relating with barcode recognition and localization in domestic and overseas, and understood the advantage of image-based methods. This work paved the way for our further research.2. We put forwarded our image processing based algorithms for localizing the position of barcode in image. Combining the concept of information entropy, we proposed the region-based gradient statistical analysis. This method has been proved to be effective and robust when dealing with images with transformations like blurring, illumination changes, scale changes, rotation, translation and pitch.3. Deep learning strategy has been introduced into barcode localization problem. Other than traditional artificial features, deep model is able to get the feature that fits the inherent statistical attributes of the dataset. We also built two datasets with 130 thousand and 94 thousand samples for training and testing PDF417 code and linear code respectively.4. With deep learning framework Caffe, we constructed a convolutional neural network with two pairs of convolution and pooling layers and two inner-product layers. Under accelerating with GPU during training and testing, we proved the effectiveness of learning algorithm and deep structure in barcode localization problem.