A Novel Convolutional Neural Network Architecture for the Classification of Binary Images

While numerous studies in the literature focus on the classification of color images using deep learning algorithms, there is a notable gap in research dedicated to the classification of binary images. Although Convolutional Neural Networks designed for binary images tend to exhibit lower performance compared to those for color images, their processing speed is significantly faster, as the input data for binary images is reduced by a factor of 24 compared with the 8-bit color images. This study aims to develop network architectures that operate with high efficiency in applications requiring only binary images, such as signature recognition, barcode reading, QR code scanning, and handwriting analysis. For this purpose, a new Bi-CNN(Binary image-CNN) network architecture was designed using existing layers. Then, a special loss function was used to improve the performance of this architecture. By integrating the classification layer called Si-CL(Signature-Classification) into Bi-CNN, a new architecture called Bi-CL-CNN emerged. Both Bi-CNN and Bi-CL-CNN were trained on two datasets. The first dataset, Shape-DU, was specifically created for testing these networks. The second dataset, MPEG-7, serves as a benchmark dataset. The performance of the trained networks is compared with three previously trained networks, namely GoogleNet, ResNet50 and DenseNet201. The empirical evaluation demonstrated that the Bi-CL-CNN network significantly outperformed the other models in both accuracy and computational speed. These findings underscore the robustness and efficiency of the proposed models in handling binary image datasets.