Deep Shallow Feature Vector Concatenation for Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification technologies using deep neural networks struggle to achieve a sufficiently high level of feature representation capability.
Innovation Solution
An image classification apparatus that extracts both low-resolution deep-layer and high-resolution shallow-layer feature vectors from input images, concatenates them, and uses these concatenated feature vectors to determine image classifications based on similarities calculated with class-specific weight matrices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If only deep-layer feature vectors are extracted for image classification, then the computational complexity is reduced, but the feature representation capability is insufficient
Solution Approach 1:
The feature extraction process is segmented into two distinct pathways: a deep-layer feature extraction pathway that captures high-level semantic information, and a shallow-layer feature extraction pathway that preserves fine-grained details. By dividing the feature extraction into multiple segments with different characteristics, the system achieves comprehensive feature representation without overwhelming computational complexity in a single pathway.
Solution Approach 2:
The patent introduces a new dimension to feature extraction by incorporating shallow-layer features alongside deep-layer features. This dimensional expansion allows the system to capture both high-level semantics and low-level details simultaneously, enhancing feature representation capability through multi-dimensional feature fusion rather than relying solely on deep-layer features.
2Reliability
If only shallow-layer feature vectors are extracted for image classification, then the fine details are preserved, but the semantic understanding is insufficient
Solution Approach 1:
The feature extraction process is segmented into two distinct pathways: a deep-layer feature extraction pathway that captures high-level semantic information, and a shallow-layer feature extraction pathway that preserves fine-grained details. By dividing the feature extraction into multiple segments with different characteristics, the system achieves comprehensive feature representation without overwhelming computational complexity in a single pathway.
Solution Approach 2:
The patent merges deep-layer feature vectors and shallow-layer feature vectors into a unified concatenated feature vector. This combination integrates the semantic understanding from deep layers with the fine detail preservation from shallow layers, creating a comprehensive feature representation that leverages the strengths of both extraction depths.
3Measurement precision
If both deep-layer and shallow-layer feature vectors are extracted and concatenated, then the feature representation capability is enhanced, but the computational cost increases
Solution Approach 1:
The patent applies partial action by selectively extracting features from specific layers rather than processing the entire feature hierarchy. By concatenating features from strategically chosen deep and shallow layers, the system achieves enhanced representation capability with reduced computational overhead compared to processing all possible feature layers, optimizing the balance between accuracy and computational cost.
Data Source
AI summary
A deep-layer feature vector extraction unit extracts a low-resolution deep-layer feature vector of an input image. A shallow-layer feature vector extraction unit extracts a high-resolution shallow-layer feature vector of the input image. A concatenation unit concatenates the deep-layer feature vector and the shallow-layer feature vector and outputs a concatenated feature vector. A similarity calculation unit retains a weight matrix of respective classes and calculates similarities from the concatenated feature vector and the weight matrix of respective classes. The shallow-layer feature vector extraction unit shares at least one convolutional layer with the deep-layer feature vector extraction unit.