Double-Column Neural Network for Image Feature Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image assessment systems rely on manually designed features that are approximations and lack a principled approach, limiting their effectiveness in capturing aesthetics-sensitive attributes, and generic features like SIFT and Fisher Vector fail to attain upper performance limits in feature-related problems due to their generic nature.
Innovation Solution
A double-column deep convolutional neural network (DCNN) is implemented to learn and classify features, using both global and local image representations, with heterogeneous inputs to automatically identify visual features and improve aesthetic quality assessment by merging layers for enhanced classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If handcrafted features are used for image assessment, then the system can capture aesthetics-sensitive attributes, but the features are manually designed with limited scope and are merely approximations of photographic or psychological rules
Solution Approach 1:
The system uses automatic feature learning through deep convolutional neural networks that learn image features autonomously from training data without manual intervention. The DCNN automatically discovers effective features through training on labeled images, eliminating the need for manual feature engineering while capturing aesthetics-sensitive attributes with higher precision and broader scope.
Solution Approach 2:
The patent transitions from fixed handcrafted features to learned features by changing the parameter representation method. Instead of using predetermined handcrafted features, the system learns optimal feature parameters automatically from data, allowing the features to adapt to different image characteristics and aesthetic criteria.
2Adaptability or versatility
If generic image features like SIFT and Fisher Vector are used, then the system can process a wide range of images, but they are unable to attain the upper performance limits in feature-related problems
Solution Approach 1:
The patent applies local quality by using a double-column DCNN architecture where one column processes global image representations and another processes local patch representations. This allows the system to maintain versatility for generic image processing while achieving high precision through specialized local feature learning that captures fine-grained details important for aesthetic assessment.
Solution Approach 2:
The system segments the image processing task into multiple components: global image processing and local patch processing. By dividing the image into patches and processing them through separate DCNN columns, the system achieves both generic applicability and high classification accuracy for aesthetic attributes.
3Measurement precision
If a double-column DCNN with merged layers is used, then classification accuracy is improved, but the device complexity increases
Solution Approach 1:
The patent merges the two DCNN columns at the fully connected layer by concatenating feature vectors from both global and local processing paths. This merging combines the benefits of global context understanding and local detail recognition, achieving high classification accuracy while maintaining a manageable network architecture through systematic integration.
Data Source
AI summary
Deep convolutional neural networks receive local and global representations of images as inputs and learn the best representation for a particular feature through multiple convolutional and fully connected layers. A double-column neural network structure receives each of the local and global representations as two heterogeneous parallel inputs to the two columns. After some layers of transformations, the two columns are merged to form the final classifier. Additionally, features may be learned in one of the fully connected layers. The features of the images may be leveraged to boost classification accuracy of other features by learning a regularized double-column neural network.