Neural Network Masking for Image Classification Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural networks used for image classification, such as in medical applications for detecting pneumonias or lung nodules from X-ray images, only provide a probability or certainty of disease presence without specifying the location within the image, making it difficult for professionals to identify the actual affected area.
Innovation Solution
A method involving the generation of a mask image to extract a target area from an input image, using a third neural network, which is then used to mask and classify the image feature map from a first neural network, allowing for classification based on the target area and generating a saliency map to highlight the affected region.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a multi-layer neural network is used for image classification, then the classification capability is improved, but the ability to locate and identify specific target areas in the image deteriorates
Solution Approach 1:
The patent divides the image processing task into two separate neural networks: one dedicated to extracting spatial location information and generating mask images, and another for classification. This segmentation allows each network to specialize in its function, with the first network preserving spatial information through mask generation and the second network performing accurate classification, thereby resolving the contradiction between classification accuracy and spatial location preservation.
Solution Approach 2:
The patent introduces mask images as an intermediary between the input image and the classification process. The mask images, generated by the first neural network, serve as a bridge that preserves spatial location information while enabling the second neural network to perform classification. This intermediary mechanism allows spatial information to be maintained and transmitted to the classification stage without being lost.
2Productivity
If the entire input image is processed for classification, then comprehensive analysis is achieved, but the ability to focus on and accurately identify target areas deteriorates
Solution Approach 1:
The patent extracts the target area from the entire input image by generating mask images that highlight only the relevant regions. The first neural network processes the full image to identify and extract target areas, creating mask images that contain only the essential information needed for classification. This extraction process improves target area detection accuracy by removing distracting background information while maintaining classification efficiency.
3Measurement precision
If a mask image generation process is added to extract target areas, then localization accuracy is improved, but the system complexity increases
Solution Approach 1:
The patent combines the mask generation and classification functions into an integrated two-network system where the outputs of the first network (mask images) directly serve as inputs to the second network. This merging approach, while adding a component, creates a cohesive system where each network's output is the other's input, allowing the system to achieve high localization accuracy through the coordinated operation of specialized networks rather than a single complex network.
Data Source
AI summary
Image classification using a generated mask image is performed by generating a mask image that extracts a target area from an input image, extracting an image feature map of the input image by inputting the input image in a first neural network including at least one image feature extracting layer, masking the image feature map by using the mask image, and classifying the input image by inputting the masked image feature map to a second neural network including at least one classification layer.