Two-Stage Image Classifier Training for Low Power Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision technologies struggle with accurately classifying objects in images with poorly localized detections, leading to high false positive rates, especially in resource-constrained camera systems with limited computational power and memory.
Innovation Solution
A two-stage training approach for deep learning neural network image classifiers, using poorly localized and accurately localized bounding boxes at different learning rates to improve classification accuracy and reduce false positives, while operating efficiently with reduced resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single-stage training approach is used with high learning rate, then training speed is fast, but classification accuracy on poorly localized detections is poor and false positives are high
Solution Approach 1:
The training process is divided into two distinct stages: first stage training with poorly localized bounding boxes at high learning rate for fast convergence, and second stage training with accurately localized bounding boxes at low learning rate for precision refinement. This segmentation resolves the contradiction by assigning different training objectives to different stages.
Solution Approach 2:
The first stage training serves as preliminary action that quickly establishes basic classification capabilities using poorly localized bounding boxes. This preliminary training enables the model to learn general object features before the second stage refines localization precision, thus achieving both speed and accuracy.
2Measurement precision
If a single-stage training approach is used with low learning rate, then classification accuracy is improved, but training time increases significantly
Solution Approach 1:
By segmenting training into two stages with different learning rates, the patent achieves high classification accuracy without requiring the entire training process to use the time-consuming low learning rate. The first stage uses high learning rate for rapid initial learning, while only the second stage uses low learning rate for precision refinement.
Solution Approach 2:
The second stage training applies partial action by focusing refinement efforts only on accurately localized bounding boxes rather than retraining on all data. This partial retraining at low learning rate achieves precision improvement without the time cost of complete retraining.
3Measurement precision
If complex classification models are used to improve accuracy on poorly localized detections, then classification performance improves, but computational resource requirements increase
Solution Approach 1:
The patent changes the parameter of bounding box quality from poor to accurate through the two-stage training process, rather than changing the model complexity. This parameter transformation achieves improved classification accuracy on poorly localized detections while maintaining the same model architecture and computational resource requirements.
4Reliability
If models are trained to reduce false positives, then detection reliability improves, but model complexity and training requirements increase
Solution Approach 1:
The training process is segmented into two stages where the second stage specifically targets false positive reduction using accurately localized bounding boxes. This segmentation allows reliability improvement through targeted refinement rather than increasing overall model complexity.
Solution Approach 2:
The second stage training uses a copy of the first stage model and refines it with low learning rate on accurately localized bounding boxes. This copying approach reduces false positives without requiring a fundamentally different or more complex model architecture.
Data Source
AI summary
Methods, systems, an apparatus, including computer programs encoded on a storage device, for training an image classifier. A method includes receiving an image that includes a depiction of an object; generating a set of poorly localized bounding boxes; and generating a set of accurately localized bounding boxes. The method includes training, at a first learning rate and using the poorly localized bounding boxes, an object classifier to classify the object; and training, at a second learning rate that is lower than the first learning rate, and using the accurately localized bounding boxes, the object classifier to classify the object. The method includes receiving a second image that includes a depiction of an object; and providing, to the trained object classifier, the second image. The method includes receiving an indication that the object classifier classified the object in the second image; and performing one or more actions.