MDig: Multi-digit Recognition using Convolutional Nerual Network on Mobile

Multi-character recognition in arbitrary photographs on mobile platform is difficult, in terms of both accuracy and real-time performance. In this paper, we focus on the domain of hand-written multi-digit recognition. Convolutional neural network (CNN) is the state-of-the-art solution for object recognition, and presents a workload that is both compute and data intensive. To reduce the workload, we train a shallow CNN offline, achieving 99.07% top-1 accuracy. And we utilize preprocessing and segmentation to reduce input image size fed into CNN. For CNN implementation on the mobile platform, we adopt and modify DeepBeliefSDK to support batching fully-connected layers. On NVIDIA SHIELD tablet, the application processes a frame and extracts 32 digits in approximately 60ms, and batching the fully-connected layers reduces the CNN runtime by another 12%.