Scale-Selective Training Data for Image Scale Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems struggle to accurately classify objects based on their scale, limiting their effectiveness in tasks like image classification and dynamic object comprehension.
Innovation Solution
The development of scale selective training data and methods to train machine learning systems to differentiate between in-scope and out-of-scope image scales, using image sensors to capture objects at varying distances and applying data augmentations to generate training data that includes desired labels based on scale.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If machine learning systems are trained to process all image scales equally, then they can handle diverse input sizes, but they fail to accurately classify objects based on their scale
Solution Approach 1:
The patent segments the image processing task by creating separate processing pathways for different scale ranges. The system divides objects into in-scope and out-of-scope categories based on scale, allowing specialized processing for each segment rather than treating all scales uniformly.
Solution Approach 2:
The patent introduces scale as an explicit dimensional parameter in the classification process. By adding scale awareness as a new dimension to the traditional classification task, the system can now distinguish objects not just by category but by their dimensional properties, resolving the contradiction between handling diversity and achieving precision.
2Measurement precision
If machine learning systems use complex training processes to improve scale awareness, then scale classification accuracy improves, but training time and computational resources increase
Solution Approach 1:
The patent applies preliminary actions by pre-processing images to annotate scale information before the main training process. Scale annotations are prepared in advance during data preparation, allowing the model to learn scale characteristics more efficiently during training rather than discovering them from scratch.
Solution Approach 2:
The patent changes key parameters of the training process by introducing scale-specific loss functions and scale-aware optimization objectives. These parameter modifications enable the model to converge faster on scale-related features by directly optimizing for scale classification accuracy rather than requiring extensive general training.
3Adaptability or versatility
If machine learning systems are trained on images with varied scales, then they can recognize objects at different sizes, but they lose the ability to make scale-selective predictions
Solution Approach 1:
The patent makes the system dynamic by enabling runtime scale parameter adjustment. The model can adapt its prediction behavior based on the detected scale of input objects, switching between in-scope and out-of-scope classification modes as needed, thus maintaining both recognition versatility and prediction selectivity.
Solution Approach 2:
The patent introduces scale annotations and scale detection mechanisms as intermediary elements between the raw image input and the final classification output. These intermediaries carry scale information through the processing pipeline, enabling the system to maintain awareness of object scale and make informed scale-selective predictions while still recognizing objects across various sizes.
Data Source
AI summary
In an aspect, the present disclosure provides a method of generating scale selective training data for use in training a machine learning system to support scale selective image classification tasks, comprising obtaining a plurality of images comprising an object of interest at a plurality of image scales; assigning a desired label to each of the plurality of images based on an image scale of the object of interest in the each image, wherein the desired label comprises an in-scope response when the image scale comprises an in-scope image scale, and generating a set of training data for use in training the machine learning system to predict a scale of the object of interest, the training data comprising the plurality of images and corresponding desired labels.