Image Feature Extraction Using DCNN and RBM Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image analysis technologies face limitations in accurately evaluating image quality and analyzing aesthetic factors due to the inadequacy of mathematical models in expressing complex features, leading to inaccurate analyses.
Innovation Solution
An image analysis method utilizing a pre-learned model based on deep convolutional neural networks (DCNN) to extract features, followed by encoding these features using Restricted Boltzmann Machines (RBM) and Support Vector Machines (SVM) to improve classification accuracy, incorporating features like quality, motion blur, and aesthetic factors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a mathematical model is designed to express aesthetic factors for image quality evaluation, then the evaluation process can be automated, but the model cannot sufficiently express various and complex features of the image leading to inaccurate analysis
Solution Approach 1:
The patent segments the image analysis process into multiple independent feature extraction modules, each targeting specific aesthetic factors (composition, color, sharpness, etc.). This segmentation allows the system to capture complex image features through multiple specialized components rather than a single mathematical model, thereby improving both automation capability and analysis accuracy.
Solution Approach 2:
The patent employs a composite approach by integrating multiple types of features (low-level features like edge and color, mid-level features like texture and shape, and high-level semantic features) into a unified evaluation framework. This composite feature representation enables the system to express various and complex image features that a single mathematical model cannot capture, resolving the contradiction between automation and precision.
2Ease of operation
If predetermined aesthetic factors are defined and a mathematical model is designed, then image quality evaluation can be performed, but the evaluation becomes inaccurate due to the model's inability to capture complex image features
Solution Approach 1:
The evaluation process is segmented into distinct stages: feature extraction, feature weighting, and quality computation. Each stage handles specific aspects of the evaluation, making the overall process systematic and manageable while improving accuracy through specialized processing at each stage.
Solution Approach 2:
The patent introduces feature weighting coefficients as intermediaries between the extracted features and the final quality evaluation. These weights act as mediators that adjust the contribution of each feature based on its importance, enabling the system to maintain operational simplicity while achieving accurate evaluations by filtering and prioritizing relevant features.
3Productivity
If a mathematical model is used to analyze aesthetic factors, then automated evaluation is achieved, but the model fails to capture the complexity and variety of actual image features
Solution Approach 1:
The patent implements dynamic feature weighting where the importance of different features is adjusted based on the specific image content and evaluation context. This dynamic adaptation allows the system to efficiently process images while maintaining versatility in capturing various image features, as the weighting mechanism automatically adapts to different image characteristics without requiring multiple fixed models.
Solution Approach 2:
The system changes parameters (feature weights, extraction thresholds, evaluation criteria) based on the input image characteristics. This parameter adaptation enables the mathematical model to handle various and complex image features efficiently, improving both productivity and adaptability by adjusting the evaluation parameters rather than redesigning the entire model for different image types.
Data Source
AI summary
An image analysis method for extracting features of an image and an apparatus for the same are disclosed. An image analysis method performed in an image analysis apparatus may comprise extracting a plurality of features for a plurality of sample images through a pre-learned model to extract features from the plurality of sample images; determining a plurality of target features representing final features to be extracted through the image analysis apparatus; encoding the plurality of features based on a probability distribution of the plurality of target features for the plurality of features; and analyzing a plurality of analysis target images based on the plurality of encoded features when the plurality of analysis target images are received.