Semantic Segmentation Network Training for Image Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In image recognition solutions, the large number of annotation dimensions and small number of training images result in poor generalization ability of trained models, leading to low accuracy.
Innovation Solution
The method involves constructing a first semantic segmentation network using an autoencoder and classifier to re-label initial labeled images, comparing them to a preset labeled result, and generating a second semantic segmentation network by training on target images with unified labels to improve recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a large number of annotation dimensions are used for training, then the model can cover more object properties, but the number of training images per annotation dimension decreases, leading to poor generalization ability
Solution Approach 1:
The patent segments the training process into two stages: first training on a subset of annotation dimensions to achieve good generalization, then gradually incorporating additional dimensions. This segmentation allows the model to learn fundamental patterns before handling more complex multi-dimensional annotations, resolving the contradiction between covering more properties and maintaining generalization ability.
Solution Approach 2:
The patent performs preliminary training on a selected subset of annotation dimensions before incorporating all dimensions. This preliminary action establishes a solid foundation of generalizable features, and subsequent fine-tuning with additional dimensions builds upon this foundation without compromising the already-learned generalization capabilities.
2Productivity
If the number of training images per annotation dimension is small, then the training process is faster and requires less computational resources, but the model's generalization ability deteriorates
Solution Approach 1:
The training process is segmented into phases with increasing complexity. The first phase uses a subset of annotation dimensions with fewer images to achieve quick convergence and good generalization. Subsequent phases gradually introduce more dimensions and images, allowing the model to build upon previously learned patterns without requiring large numbers of images for each individual dimension.
Solution Approach 2:
The patent dynamically adjusts training parameters including the subset of annotation dimensions used, learning rates, and optimization strategies based on training progress. This adaptive parameter adjustment allows the model to achieve good generalization with fewer images per dimension by optimizing the training process itself rather than relying solely on increasing data quantity.
Data Source
AI summary
An image recognition method applied to an electronic device is provided. The method includes constructing a first semantic segmentation network. In response that an initial labeled result of one of a plurality of initial labeled images does not match a preset labeled result, a target image corresponding to the one of the plurality of initial labeled images and a target labeled result of the target image are obtained. A second semantic segmentation network is obtained by training the first semantic segmentation network based on a plurality of the target images and the target labeled result of each target image, and a labeled result of an image to be recognized is obtained by inputting the image to be recognized into the second semantic segmentation network.