Kernel Pooling Layer for Finer Picture Recognition Granularity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current convolutional neural network models for picture recognition lack granularity and accuracy due to limited pooling operations, resulting in undesirable recognition performance when classifying large datasets.
Innovation Solution
Incorporating a kernel pooling layer that upgrades the dimensionality of the channel from one dimension to multiple dimensions within the convolutional neural network model, enhancing the representation performance and enabling finer granularity in picture classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional pooling operations (average pooling, maximum pooling, bilinear pooling) are used in the convolutional neural network model, then the model structure remains simple and computationally efficient, but the picture recognition accuracy and granularity are insufficient
Solution Approach 1:
The patent introduces kernel pooling layer that upgrades the dimensionality of channel from one dimension to multiple dimensions. This dimensional transformation enables the model to capture more nuanced features and achieve finer granularity in picture classification, directly resolving the accuracy-granularity limitation of traditional pooling operations.
Solution Approach 2:
The patent segments the pooling operation into multiple kernel pooling layers, each with different kernel sizes and configurations. This segmentation allows the model to process features at multiple scales and levels of abstraction, improving recognition accuracy without requiring a complete redesign of the overall model structure.
2Loss of information
If traditional pooling operations are used, then the computational complexity remains low, but the representation performance and granularity of picture recognition are limited
Solution Approach 1:
By transforming the channel dimensionality from 1D to multi-dimensional in the kernel pooling layer, the patent enriches feature representation without requiring proportional increases in computational power. The dimensional transformation naturally captures more feature interactions and relationships.
Solution Approach 2:
The patent adjusts key parameters in the kernel pooling layer, including kernel size, stride, and padding configurations, to optimize the balance between feature representation quality and computational efficiency. These parameter changes enable better information retention with manageable computational costs.
Data Source
AI summary
The present disclosure provides a picture recognition method and apparatus, a computer device and a computer-readable medium. The method comprises: obtaining a to-be-recognized picture; according to the to-be-recognized picture and a pre-trained picture recognition model, acquiring a predicted class tag of the to-be-recognized picture; the pre-trained picture recognition model employs a convolutional neural network model, and the convolutional neural network model comprises at least one layer of kernel pooling layer upgrading a dimensionality of a channel from one dimension to multiple dimensions; recognizing a class of the to-be-recognized picture according to the predicted class tag of the to-be-recognized picture. By recognizing the to-be-recognized picture by using the picture recognition model having the kernel pooling layer, the present disclosure can achieve recognition of picture classes with finer granularities and effectively improve the picture recognition accuracy and efficiency.