Introduction
Modern imaging sensors frequently capture high-resolution data (10-20 megapixels) containing critical fine-grained details such as faces or license plates. However, standard computer vision pipelines often downscale these images to low resolutions (e.g., 224x224) to fit within the input constraints of Convolutional Neural Networks (CNNs). In scenarios involving remote processing over bandwidth-limited networks—such as satellite links, drone operations, or disaster response—transmitting full high-resolution images is often impossible. Conversely, aggressive JPEG compression or downscaling destroys the semantic details required for tasks like Optical Character Recognition (OCR).
Active Information Seeking
To bridge this gap, this research proposes an "active information seeking" framework. Instead of transmitting the entire scene, the system identifies regions of interest in a low-resolution preview and requests only those specific crops at high resolution. This approach treats bandwidth optimization as a semantic cropping problem, prioritizing pixels that contribute to the final recognition task.
Hierarchical Recognition Models
The authors focus on a hierarchical relationship: detecting cars (parent objects) to recognize license plates (sub-objects). Two primary architectures are explored:
- Recursive Fast-RCNN: This model applies a standard detector iteratively. It first identifies cars in a low-resolution image, requests high-resolution crops of those cars, and then runs a second detector to find license plates within those crops for OCR.
- Multi-stage Network: This novel architecture optimizes bandwidth further by predicting both the car and the license plate locations simultaneously from the low-resolution image. By estimating the sub-object's coordinates early, the system can bypass the "car crop" stage and request only the tiny license plate pixels directly.
Experimental Setup and Dataset
A new dataset of 500 high-resolution images was developed to benchmark this approach. Unlike existing datasets that focus on isolated objects, this collection includes full scenes with manually annotated bounding boxes for both cars and plates, along with transcriptions and difficulty ratings. This allows for a holistic evaluation of the detection-to-recognition pipeline.
Key Results and Trade-offs
The proposed framework achieves up to an order of magnitude in bandwidth savings compared to standard JPEG compression baselines while maintaining high accuracy.
- Efficiency: The Multi-stage Network is the most bandwidth-efficient, reducing transmission size by an additional 20% over the Recursive model (averaging ~0.058MB vs 0.073MB per image).
- Accuracy: The Recursive model maintains 64% character accuracy, nearly matching the performance of processing the original uncompressed image. The Multi-stage model sees a slight 5% drop in accuracy, primarily due to the difficulty of precisely localizing tiny sub-objects from low-resolution inputs.
Conclusion
By leveraging the semantic hierarchy of objects, "active information seeking" allows for high-fidelity recognition in environments where bandwidth is a critical bottleneck. Future work aims to combine this semantic cropping with traditional compression techniques and improve sub-object localization through larger-scale training.
Summary updated