A Compact n x 2 Descriptor Matrix for Efficient Binary Image Representation in Neural Networks
El Kharrachi Khayya, Ahmed El Oirrak, Toufik Datsi · 2025
This article proposes a novel way of representing binary images in a compact and operationally efficient fashion for boosting the performance of neural networks on binary classification tasks. While traditional Convolutional Neural Networks (CNNs) extract spatial features with convolutional layers that can be computationally heavy for large datasets, they exhaust the spatial information in an image. This is circumvented by our approach, which transforms each binary image into a descriptor matrix of size “nx2”, where “n” denotes number of rows in the image, and each row in the descriptor encodes the pixel count and transition information. However, this feature-based representation greatly reduces the input size, making the processing faster and reducing the computational overhead. It is then used as an input in a fully connected neural network leading to competitive results at a fraction of the computational cost. We experiment with MNIST dataset and show how this work can achieve a comparable accuracy to traditional CNN based solutions at a significantly lower parameter and computational costs.