Image Classification Feature-Matrix Rearrangement for Faster Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high computational demands in fully connected layers of image classification networks lead to slow processing speeds, necessitating improvements in efficiency while maintaining classification accuracy.
Innovation Solution
The method involves training a classification model using a convolutional neural network, adjusting it based on prediction accuracy, and optimizing the dimensionality and rearrangement of feature matrices to enhance the fully connected layer operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fully connected layer is used in the image classification network, then the classification accuracy can be maintained, but the computation time increases significantly
Solution Approach 1:
The patent extracts and removes the fully connected layer from the image classification network, replacing it with a classification head that operates on feature maps from the convolutional base. This extraction eliminates the computationally expensive fully connected operations while preserving classification functionality through alternative mechanisms such as global average pooling followed by classification layers.
2Reliability
If the fully connected layer computations are performed, then comprehensive feature analysis is achieved, but the processing speed decreases
Solution Approach 1:
The patent replaces the mechanical computation system of fully connected layers with a more efficient approach using global average pooling and lightweight classification layers. This substitution maintains the ability to analyze features comprehensively while dramatically reducing the computational burden and improving processing speed.
Data Source
AI summary
An image classification method applied to an electronic device is provided. The method includes obtaining prediction results of a plurality of test images. A plurality of target images is determined according to the prediction results, and a prediction accuracy rate of a first classification model is calculated based on the plurality of target images. A second classification model is obtained according to the plurality of target images. An initial feature matrix is obtained using the second classification model. An input feature matrix is obtained by performing a dimension raising process on an initial feature matrix. Once a target weight matrix is obtained based on an initial weight matrix, and a target vector is generated according to the target weight matrix and the input feature matrix, a classification result of an image to be classified is obtained by inputting the target vector into a classification layer.