OCR Studio Presented 40× AI Model Compression Technology at ACM SenSys 2026

OCR Studio, a developer of optical character recognition technologies, debuted at SenSys 2026 with a new architectural block for computer vision neural networks that reduces model size by more than 40 times without sacrificing accuracy. The solution is aimed at industries requiring real-time on-device processing, including fintech, digital identity verification, logistics, and enterprise document processing.

As modern AI systems continue to improve in accuracy, they become larger and more computationally expensive. This limits their deployment on edge devices, where memory, processing power, latency, and energy consumption are critical constraints. For businesses integrating AI into real-world products and services compact models are becoming important for reducing infrastructure costs, enabling real-time and on-device processing. This is especially relevant for computer vision tasks requiring full-image context understanding, where attempts to decrease model size often lead to lower recognition quality and compromised reliability in production scenarios.

To address this problem, OCR Studio scientists developed a new Infinite Impulse Response (IIR) Block – a neural network module based on learnable recursive filters. Unlike conventional convolutions operating within a fixed local window, the IIR Block propagates information across the entire image, effectively expanding the receptive field without increasing network depth. This allows developers to replace dozens of layers with a compact module while preserving model quality.

The researchers integrated the IIR Blocks into a standard U-Net architecture used for document image enhancement and OCR preprocessing of low-quality documents. The resulting model contains only 49 thousand parameters compared to approximately 2 million parameters in the baseline model, making it 40.8 times smaller. At the same time, the compact network maintained competitive performance on the DIBCO benchmarks.

During SenSys 2026 OCR Studio researchers also presented a paper focused on real-time recognition of passport machine-readable zone (MRZ) on augmented reality glasses. The scientists introduced an ultra-lightweight end-to-end recognition pipeline capable of processing MRZ data directly on AR devices without transferring images to external servers. The solution was developed specifically for devices operating under strict hardware limitations, including low battery capacity, limited processor performance, and poor-quality cameras that often struggle with unstable lightning conditions, motion blur, and challenging capture scenarios.

ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems (SenSys 2026) is a leading international forum for research on systems and algorithms for embedded, energy-constrained, and sensing devices. The conference brings together researchers and industry experts to discuss the latest advances in making intelligent systems more practical, efficient, and deployable in real-world environments.