Image Processing Device Using Knowledge Distillation for Data Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing techniques that reduce data size while maintaining accuracy often result in decreased recognition accuracy of post-conversion images, as they fail to effectively inherit knowledge from the original image, leading to altered ranking and classification of image categories.
Innovation Solution
An image processing device that incorporates a knowledge distillation mechanism to maintain the accuracy of post-conversion images by learning from the original image's knowledge, ensuring that the post-conversion image retains the same trend in knowledge as the original image, while reducing data size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If image conversion is performed to reduce data size, then code amount is reduced, but recognition accuracy of post-conversion images decreases
Solution Approach 1:
A knowledge distillation model is introduced as an intermediary between the original image and the post-conversion image. This model learns the knowledge (classification patterns and relationships) from the original image and transfers it to the post-conversion image, enabling the compressed image to maintain recognition accuracy comparable to the original image while achieving significant data size reduction
Solution Approach 2:
The knowledge distillation model performs preliminary learning on the original image before conversion. By pre-learning the classification knowledge and relationships from the original image, the model ensures that this knowledge is preserved and transferred to the post-conversion image, preventing accuracy loss during the compression process
2Quantity of substance
If conventional image conversion is used, then code amount is reduced, but knowledge and category ranking are altered
Solution Approach 1:
The knowledge distillation model serves as a mediator that captures and preserves the classification knowledge from the original image. It learns the relationships between categories and objects in the original image and transfers this knowledge to the post-conversion image, ensuring that category rankings and semantic information are maintained despite compression
Solution Approach 2:
The knowledge distillation process creates a copy of the classification knowledge from the original image. By copying the essential knowledge patterns, category relationships, and ranking information, the system ensures that the post-conversion image retains the same knowledge structure as the original, preventing information loss during compression
Data Source
AI summary
An image processing device includes an image processing unit configured to execute image processing on an image based on an input image, and output a result of the image processing. The input image is a post-conversion image obtained by performing image conversion on an original image. The image conversion includes image conversion for further reducing a data size of the original image while maintaining a feature acquired from the original image and related to an object similar to a subject captured in the original image and maintaining processing accuracy of the image processing.