Composite Neural Network Hierarchical Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to accurately identify both primary segments of objects and subsegments within those objects in images, often requiring separate and larger neural networks for each detection, leading to increased computational resources and reduced accuracy due to the inability to learn object hierarchies.
Innovation Solution
The training of composite neural networks, where a primary part is trained to detect primary segments and its activations are used to train secondary parts to detect subsegments, allowing for a smaller network size and improved accuracy by learning hierarchical relationships between objects.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If separate neural networks are used for detecting primary segments and subsegments, then detection coverage is improved, but device complexity and computational resources increase
Solution Approach 1:
The patent combines multiple detection functions into a single composite neural network that processes images to detect both primary segments and subsegments hierarchically. The network integrates primary segment detection and subsegment detection capabilities within one unified architecture, reducing the need for separate networks while maintaining comprehensive detection coverage.
Solution Approach 2:
The composite neural network is divided into distinct functional parts: a primary part that detects primary segments and a secondary part that detects subsegments within those primary segments. This segmentation allows each part to specialize in specific detection tasks while working together within the unified network structure, improving efficiency without requiring completely separate networks.
2Measurement precision
If larger neural networks are used for each detection task, then detection accuracy is improved, but computational resources and training time increase
Solution Approach 1:
The primary part of the composite neural network performs preliminary detection of primary segments first, establishing bounding boxes and regions of interest. This preliminary action allows the secondary part to focus computational resources only on detecting subsegments within these identified regions, rather than searching the entire image, thereby reducing overall computational requirements while maintaining detection accuracy.
Solution Approach 2:
The network applies different processing quality and computational effort to different regions of the image based on the presence of primary segments. Areas containing primary segments receive more detailed processing for subsegment detection, while areas without primary segments receive minimal processing. This local quality approach optimizes computational resource allocation according to actual detection needs.
3Ease of manufacture
If separate neural networks are trained independently, then training flexibility is improved, but training time and loss of time increase
Solution Approach 1:
The primary part of the composite neural network is trained first to detect primary segments, establishing a foundation for subsequent subsegment detection. This preliminary training action allows the secondary part to be trained afterward using the outputs and activations from the primary part, enabling a staged training approach that reduces overall training time compared to training completely separate networks independently.
Solution Approach 2:
The primary part serves as an intermediary between the input image and the secondary part. Its activations and outputs act as intermediate representations that the secondary part builds upon during training. This intermediary relationship enables efficient knowledge transfer and reduces the training burden on the secondary part, as it doesn't need to learn from raw images but rather from pre-processed features.
Data Source
AI summary
A method of training a machine learning model includes training a primary part of a composite neural network to identify a primary segment of objects in a training image, freezing the primary part of the composite neural network after training the primary part of the composite neural network, and after freezing the primary part of the composite neural network, training, using activations of the primary part of the composite neural network, a secondary part of the composite neural network to identify a first subsegment of objects or a feature of the first subsegment of objects in the training image. The first subsegment of objects is a subset of the primary segment of objects.