Digital Halftone Classification using Simplified CNN and Stochastic Statistics

Digital halftone technique is widely utilized in various printing applications to generate high quality printing output. The halftone classification is a fundamental requirement for the optimal reconstruction of binary printed images during inverse half toning and image authentication. In this study, a new classification scheme based on convolutional neural networks (CCN) is proposed. For effective training, a new approach exploiting the properties of stochastic halftone parameters is proposed to obtain the favorable halftone patch. The CNN network has been optimized to perform classification with minimal layers, and the model is trained using cross entropy, and make use of softmax classification layer. During testing, the sub-halftone patches are classified individually and then majority voting strategy is performed to identify the actual halftone class. For performance evaluation, a deep learning halftone database has been developed for the 24 halftone varieties which also includes the latest multitone images. From experimental result, it has been validated that the proposed architecture achieves average correct classification rate (ACCR) of 100% for several halftone class, and showed superior performance in comparison with state-of-the-art methods.