Segmented Image Classification Without Global Average Pooling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image classification methods using neural network models require significant memory and high power consumption due to the complexity of global average pooling layers.
Innovation Solution
Replace global average pooling layers with convolutional and average pooling layers in classification models, perform image segmentation and normalization, and generate target vectors based on pixel average values to reduce model parameters and improve feature extraction speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If global average pooling layers are used in neural network models for image classification, then classification capability is maintained, but memory requirements increase and power consumption increases
Solution Approach 1:
The patent extracts and removes the global average pooling layers from the neural network model architecture. By eliminating these computationally intensive layers, the model reduces memory requirements while maintaining classification capability through alternative feature extraction mechanisms that operate on segmented image regions rather than global pooling.
Solution Approach 2:
The patent employs copying by creating multiple simplified models with different depths (first model, second model, third model) that replicate the classification functionality of the original complex model but with reduced computational complexity. These copied models process segmented image regions independently, achieving the same classification result with lower memory and power consumption.
2Reliability
If global average pooling layers are used in neural network models for image classification, then classification capability is maintained, but power consumption increases
Solution Approach 1:
The patent extracts and removes the global average pooling layers from the neural network model architecture. By eliminating these computationally intensive layers, the model reduces memory requirements while maintaining classification capability through alternative feature extraction mechanisms that operate on segmented image regions rather than global pooling.
Solution Approach 2:
The patent employs copying by creating multiple simplified models with different depths (first model, second model, third model) that replicate the classification functionality of the original complex model but with reduced computational complexity. These copied models process segmented image regions independently, achieving the same classification result with lower memory and power consumption.
3Reliability
If global average pooling layers are used in neural network models, then classification is performed, but hardware resource requirements increase
Solution Approach 1:
The patent applies segmentation by dividing the image into multiple regions and processing each region independently through separate convolutional operations. Instead of using global average pooling on the entire image, the model segments the image and performs localized feature extraction, reducing hardware resource requirements while maintaining classification accuracy through region-based analysis.
Solution Approach 2:
The patent employs copying by creating multiple simplified models with different depths (first model, second model, third model) that replicate the classification functionality of the original complex model but with reduced computational complexity. These copied models process segmented image regions independently, achieving the same classification result with lower memory and power consumption.
Data Source
AI summary
A method for identifying images of any kind of object, applied in an electronic device, obtains images to be classified, and obtains classification models and target models. Image segmentation on the images is carried out and target images are obtained; target images are input into each of the input layers and target feature maps are generated; target values according to size of each target feature map are generated; a convolution operation on the target feature maps generates feature values and a target vector of each target model is generated according to average pooling layers, feature values, and target values. The target vector is input into classification layers of a target model and label categories and probability values of categories obtained, enabling classification of images according to the label categories and the probability values of categories.