Indoor Scene Classification via Iterative Local Observation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional indoor scene recognition methods are not accurate due to reliance on low-level geometric information and limited extraction of high-level semantic information, and the convolutional neural network-based methods consume significant resources and fail to grasp the fusion of overall and local information, resulting in low classification rates.
Innovation Solution
A classification method that processes an input scene picture by obtaining a current local observation area, generating a feature vector, and iteratively refining the classification prediction until it meets a predetermined condition, using a predetermined observation area positioning model to adjust the local observation area and improve feature vector comprehensiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the spatial pyramid method is used for scene recognition, then the geometric information is processed, but the high-level semantic information extraction is limited reducing classification accuracy
Solution Approach 1:
The patent merges multiple feature extraction methods including spatial pyramid pooling, high-level semantic feature extraction through multiple convolutional layers, and color histogram analysis. This combination allows the system to capture both geometric low-level features and semantic high-level features simultaneously, resolving the contradiction between geometric processing and semantic information extraction.
2Measurement precision
If the convolutional neural network method is used for object detection and classification, then the object recognition rate reaches 94%, but the training process consumes a lot of resources and the indoor scene classification rate is only 69%
Solution Approach 1:
The patent segments the scene recognition task into multiple independent feature extraction components: spatial pyramid pooling for geometric structure, color histogram for visual appearance, and selective convolutional feature extraction for semantic information. Each component processes specific aspects of the image independently, then fuses results. This segmentation reduces training resource consumption compared to training a full CNN while maintaining effective classification capability.
Solution Approach 2:
The patent applies local quality by selectively extracting features from specific regions and using different processing strategies for different parts of the image. The observation area positioning model focuses computational resources on relevant local regions rather than processing the entire image uniformly, reducing overall computational complexity and resource consumption while maintaining high recognition accuracy.
3Measurement precision
If the convolutional neural network extracts features directly, then the object features are obtained, but the fusion of overall information and local information cannot be grasped reducing scene classification accuracy
Solution Approach 1:
The patent implements a nested structure where spatial pyramid pooling operates at multiple scales (nested levels of granularity), color histograms are computed at different resolutions, and convolutional features are extracted at multiple depths. This nested multi-scale approach allows the system to capture both global scene context and local object details simultaneously, achieving effective fusion of overall and local information for improved scene classification accuracy.
Data Source
AI summary
The present disclosure provides a classification method and a classification device of an indoor scene. The classification method includes steps: receiving an input scene picture to be classified; obtaining a current local observation area from the scene picture to be classified according to a observation area positioning model; processing image information of the current local observation area to obtain a feature vector; obtaining a classification prediction result according to the feature vector; determining Whether the classification prediction result satisfies a predetermined scene picture classification condition; if not, obtaining a next local observation area from the scene picture to be classified and setting the next local observation area to be the current local observation area, then skipping to the step of processing the image information of the current local observation area; if yes, obtaining a classification label of the scene picture to be classified.