Memory Efficient Local Features Descriptor for Identity Document Detection on Mobile and Embedded Devices

In this paper, we propose a data-driven approach to training a memory-efficient local feature descriptor for identity documents location and classification on mobile and embedded devices. The proposed algorithm for retrieving a dataset of patches is based on the specifics of document detection in smartphone camera-captured images with a template matching approach. The retrieved dataset of patches relevant to the domain, which includes splits for features training, features selection, and testing, is made public. We train a binary descriptor using the retrieved dataset of patches, each bit of the descriptor relies on a single computationally-efficient feature. To estimate the influence of different feature spaces on the descriptor performance, we perform descriptor training experiments using gradient-based and intensity-based features. Extensive experiments in identity document location and classification benchmarks showed that the resulting 128 and 192-bit descriptors which use gradient-based features outperformed a state-of-the-art 512-bit BEBLID descriptor for arbitrary keypoints matching in all cases except the cases of extreme projective distortions, being significantly more efficient in cases of low lighting. The 64-bit gradient-based descriptor obtained within the approach showed better quality than 128 and 256-bit BinBoost descriptors in scanned document images. To evaluate the influence of the descriptor size on the matching speed, we propose a model based on the required number of processor instructions for computing the Hamming distance between a pair of descriptors on various energy-efficient processor architectures.