CNN Image Classification via Feature Map Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, such as convolutional neural networks, face challenges in accurately classifying images due to overlapping features between different anomaly categories, leading to poor performance in recognizing less common anomalies or those similar to other classes, especially when training data is imbalanced.
Innovation Solution
The method involves generating a feature map from the model's layers to identify overlapping areas between image categories, splitting the categories into subcategories, and retraining the model using these subcategories to improve classification accuracy by balancing the training data and reducing overlap.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a CNN is trained on a standard imbalanced training data set, then the model can recognize common anomalies, but it performs poorly on less common anomalies or anomalies that appear similar to other classes
Solution Approach 1:
The patent segments the training data by identifying overlapping regions in the feature map and creating separate sub-datasets for overlapping and non-overlapping images. This segmentation allows the model to learn distinct features for anomalies that are similar to other classes versus those that are distinct, thereby improving reliability for classifying less common anomalies without requiring more training images.
2Quantity of substance
If training data includes all detected anomalies over time, then the data set is comprehensive, but the data becomes imbalanced with too many common anomalies and too few less common anomalies
Solution Approach 1:
The patent segments the comprehensive training data into overlapping and non-overlapping subcategories based on feature map analysis. By creating balanced sub-datasets with equal numbers of images from each category, the method maintains complete anomaly coverage while achieving classification balance across different anomaly frequencies.
Solution Approach 2:
The patent changes the parameter of data distribution by creating artificially balanced sub-datasets. Instead of using the naturally imbalanced distribution from production data, the method adjusts the parameters by selecting equal numbers of images from overlapping and non-overlapping categories, thereby achieving manufacturing precision in classification across all anomaly types.
3Ease of operation
If categories are kept broad to simplify classification, then the model is easier to operate, but images with overlapping features are misclassified
Solution Approach 1:
The patent segments broad categories into subcategories based on whether images overlap with other categories in the feature map. This segmentation creates a more detailed classification structure that maintains ease of operation through systematic subcategorization while improving measurement precision by specifically addressing overlapping images with dedicated training subsets.
Data Source
AI summary
The present disclosure relates to training a machine learning model to classify images. An example method generally includes receiving a training data set including images in a first category and images in a second category. A convolutional neural network (CNN) is trained using the training data set, and a feature map is generated from layers of the CNN based on features of images in the training data set. A first area in the feature map including images in the first category and a second area in the feature map where images in the first category overlap with images in the second category are identified. The first category is split into a first subcategory corresponding to the first area and a second subcategory corresponding to the second area. The CNN is retrained based on the images in the first subcategory, images in the second subcategory, and images in the second category.