Directional Image Upsampling for Resource-Limited Devices

The patent segments the image upsampling process into two distinct stages: (i) an upsampling stage that generates intermediate pixels at target locations, and (ii) an adaptive sharpening stage that selectively enhances edges and features. This segmentation allows each stage to be optimized independently, with the upsampling stage using efficient algorithms and the sharpening stage using a compact neural network, thereby reducing overall processing time while maintaining output quality.

The patent extracts and isolates the most computationally intensive aspect of neural network-based upsampling (the adaptive filtering/sharpening operation) into a separate, optimized stage. By taking out this critical function and implementing it as a dedicated adaptive sharpening stage with pre-computed kernels and directionality parameters, the system reduces the computational burden on the main processing pipeline while preserving the quality benefits.

The patent segments the processing into an efficient upsampling stage and a compact adaptive sharpening stage, allowing the majority of pixels to be processed using low-power algorithms while applying neural network-based enhancement only where needed. This segmentation dramatically reduces overall power consumption compared to applying a full neural network to all pixels.

The patent applies adaptive sharpening selectively based on local image characteristics such as edge directionality and gradient magnitude. By computing directionality parameters and applying enhanced filtering only in regions that benefit from it (e.g., edges and high-frequency regions), the system maintains high output quality while minimizing power consumption by avoiding unnecessary computations in smooth regions.

The patent extracts the sharpening function as a separate post-processing stage that operates on the output of conventional upsampling methods. By taking out this critical quality-enhancement function and implementing it as an adaptive sharpening stage with directionality-aware filtering, the system preserves the speed and efficiency of conventional methods while adding the quality improvement that was previously only available through slow full neural network processing.

The patent applies adaptive sharpening selectively based on local image characteristics such as edge directionality and gradient magnitude. By computing directionality parameters and applying enhanced filtering only in regions that benefit from it (e.g., edges and high-frequency regions), the system maintains high output quality while minimizing additional processing overhead, thus resolving the contradiction between speed and quality.

The patent performs preliminary computation of directionality parameters (such as gradient directions and edge orientations) before applying the adaptive sharpening filters. By pre-computing these parameters and organizing them in a structured manner, the system reduces the complexity of the subsequent filtering stage, as the adaptive kernels can be selected and applied more efficiently without re-computing directional information.

The patent changes the parameters of the filtering operation dynamically based on local image characteristics, specifically using directionality parameters to adapt the orientation and strength of sharpening filters. By parameterizing the filtering process in terms of directionality and gradient magnitude, the system achieves adaptive quality enhancement while maintaining manageable computational complexity through efficient parameter computation and kernel selection.