Vision-Language Model Image Classification Using Decision Graph
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The pace of performance improvement in artificial intelligence, particularly in image classification tasks using vision-language models, has been slow, and there is a challenge in classifying images into complex classes with existing technologies.
Innovation Solution
A vision-language model-based image classification method that involves entering a data list of images, specifying the classification, and using a prompt with words classified into at least one class. This method reads a graph defining the classification order, selects nodes, reads class information, and classifies images based on the prompt and class information, returning a classification result if the node is a leaf node.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional vision-language models are used for image classification, then the system structure remains simple, but the classification accuracy for complex classes is insufficient
Solution Approach 1:
The patent segments the image classification task into multiple hierarchical levels using a decision graph structure. Instead of performing single-shot classification, the system divides complex classification into sequential steps where each node in the graph represents a classification stage, progressively narrowing down categories from broad to specific, thereby improving accuracy without requiring a completely complex system redesign
Solution Approach 2:
The patent applies preliminary action by pre-defining the decision graph structure and class hierarchies before actual classification occurs. The system prepares classification prompts and organizes class relationships in advance, allowing the vision-language model to operate more efficiently during inference by following predetermined classification paths rather than exploring all possibilities simultaneously
2Productivity
If zero-shot method is used with vision-language models, then prior learning time is reduced, but performance improvement pace is slow
Solution Approach 1:
The patent segments the classification process into multiple stages represented by nodes in a decision graph. Each stage performs a focused classification task on a subset of classes rather than attempting to classify all classes simultaneously. This segmentation allows the zero-shot method to maintain speed while improving accuracy through progressive refinement at each classification stage
Solution Approach 2:
The patent introduces dynamic adaptability by allowing the system to traverse different paths through the decision graph based on the specific image being classified. The classification process dynamically adjusts which prompts are used and which class hierarchies are explored, optimizing the balance between speed and accuracy for each individual classification task
3Adaptability or versatility
If complex classes are included in classification, then classification coverage is improved, but false-positive results increase
Solution Approach 1:
The patent segments complex class classification into hierarchical levels where broader categories are classified first, then progressively refined into more specific sub-categories. This segmentation reduces false positives by ensuring that each classification decision is made in the context of previously established broader categories, providing multiple levels of verification
Solution Approach 2:
The patent implements feedback mechanisms through the decision graph structure where classification results from one node inform the selection of subsequent nodes and prompts. The system uses the output of each classification stage as feedback to guide the next stage, allowing incorrect classifications to be caught and corrected at subsequent levels, thereby improving result reliability
Data Source
AI summary
The present disclosure relates to the use of artificial intelligence and, more particularly, to an efficient use of a vision-language model. A vision-language model-based image classification method according to one embodiment of the present disclosure comprises entering a data list including at least one image, entering a prompt specifying classification of an image, extracting a word classified into at least one class from the prompt, reading a graph defining a classification order of at least one class, selecting a node of the graph, reading class information corresponding to the node, classifying an image based on the prompt and class information, and returning a classification result if the node is a leaf node.