Neural Image Processing With Object-Specific Models and Internal Memory
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies using artificial neural networks consume high power due to frequent memory access for kernel reading, hindering their application in devices with limited power resources.
Innovation Solution
Implement a neural processing unit (NPU) with internal memory to store models, allowing for faster access and reduced energy consumption by processing images using multiple independent neural network models tailored to specific objects, improving image quality through selective application of models based on classification results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If convolutional neural network operations are performed using general-purpose processors with external memory, then image processing capability is achieved, but power consumption increases due to frequent kernel reading from memory
Solution Approach 1:
The system segments the neural network processing into two distinct parts: a classification neural network for object identification and a processing neural network for image quality enhancement. This segmentation allows each network to be optimized for its specific function and enables selective execution based on classification results, reducing unnecessary computations and power consumption while maintaining processing speed for critical tasks.
Solution Approach 2:
The classification neural network performs preliminary object identification and classification before the processing neural network applies image quality enhancement. This preliminary action enables the system to determine which processing operations are necessary based on the classified objects, avoiding unnecessary processing steps and reducing overall power consumption while maintaining efficient processing speed for relevant operations.
2Manufacturing precision
If multiple independent neural network models are used for different objects, then image quality improvement is enhanced, but device complexity increases
Solution Approach 1:
The system dynamically selects and executes processing neural networks based on the classification results from the classification neural network. Instead of continuously running all processing models, the system activates only the relevant processing networks corresponding to the classified objects in the current image. This dynamic approach maintains high image quality for relevant objects while reducing device complexity and power consumption by deactivating unnecessary processing networks.
Solution Approach 2:
Different processing neural networks are applied to different regions or objects in the image based on their classification. Each processing network is specialized for enhancing specific types of objects (e.g., portraits, landscapes, food), providing optimal local quality enhancement for each object type while avoiding the complexity of applying all processing networks uniformly across the entire image.
3Productivity
If classification results are used to selectively apply processing models, then processing efficiency is improved, but control complexity increases
Solution Approach 1:
The classification neural network serves a universal function by providing classification results that guide multiple different processing neural networks. Instead of creating separate control mechanisms for each processing network, the single classification output is universally used to route to the appropriate processing networks. This multi-functional approach improves processing efficiency by enabling selective execution while avoiding the complexity of multiple independent control systems.
Data Source
AI summary
An image processing method includes receiving an image including an object; classifying at least one object in the image using a first model on the basis of an artificial neural network configured to classify the at least one object by inputting the image; and obtaining an image having improved quality according to the at least one object by inputting the image in which the at least one object is classified by using at least one model among a plurality of second models on the basis of an artificial neural network configured to output a specialized processing applied image according to a particular object by inputting the received image.