Classification Model Optimization for Image Feature Expressiveness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Classification models used for image feature acquisition have reduced expressiveness when dealing with a large number of product image classes, leading to inaccurate classification results.
Innovation Solution
The method involves training a classification model using preset classes of training images, determining nonsimilar image pairs through verification images, and optimizing the model based on both similar and nonsimilar image pairs to improve feature expressiveness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a classification model is trained using preset classes of training images, then the model can perform classification, but the expressiveness of extracted image features is reduced when the quantity of classes is relatively large
Solution Approach 1:
The patent segments the feature extraction process into two distinct stages: initial feature extraction using preset classes, and subsequent optimization using constructed similar and nonsimilar image pairs. This segmentation allows the model to first learn basic classification and then refine feature expressiveness through targeted optimization, resolving the contradiction between maintaining classification capability and improving feature quality for large-scale classes.
Solution Approach 2:
The patent performs preliminary action by constructing similar and nonsimilar image pairs in advance before the optimization training phase. These pre-constructed pairs serve as targeted training data that addresses the specific problem of reduced feature expressiveness. By preparing this specialized training data beforehand, the model can efficiently optimize its feature extraction capability without requiring extensive additional data collection during the optimization phase.
2Adaptability or versatility
If the quantity of classes for product images is increased, then more product categories can be classified, but the classification model accuracy is reduced
Solution Approach 1:
The patent divides the training process into two segments: initial training with preset classes to establish basic categorization capability across many product types, and subsequent optimization training using constructed similar and nonsimilar image pairs to refine accuracy. This segmentation enables the model to handle a large number of classes while maintaining high classification accuracy through targeted refinement.
Solution Approach 2:
The patent changes the training parameters by transitioning from standard classification training to optimization training using specially constructed image pairs. This parameter change involves modifying the loss function and training data structure to focus on distinguishing between similar classes, thereby improving accuracy without sacrificing the ability to handle diverse product categories.
3Ease of manufacture
If standard classification model training is used, then the training process is simple, but the extracted image features have reduced expressiveness for different classes
Solution Approach 1:
The patent segments the training process into an easy initial phase using preset classes and a refinement phase using constructed image pairs. This segmentation maintains ease of implementation by building upon standard training procedures while adding a targeted optimization step that improves feature expressiveness without requiring complete redesign of the training pipeline.
Solution Approach 2:
The patent performs preliminary construction of similar and nonsimilar image pairs before optimization training. This preliminary action prepares the necessary training data in advance, making the optimization process straightforward to implement. The pre-constructed pairs enable the model to efficiently improve feature expressiveness without complex real-time processing during training.
Data Source
AI summary
The present application provides an image feature acquisition method and a corresponding apparatus. According to an example of the method, a classification model may be trained by using preset classes of training images, and similar image pairs may be determined based on the training images; classification results from the classification model are tested by using verification images to determine nonsimilar image pairs; and the classification model is optimized based on the similar image pairs and the nonsimilar image pairs. In this way, the optimized classification model may be used to acquire image features.