Multiscale Feature Representations for Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multiscale representation methods for object recognition in computer vision are computationally costly and inefficient, as they require multiple computations at different scales, and intermediate-level features in CNNs are less discriminative.
Innovation Solution
A method involving a processor that downsamples input data and runs a first convolutional neural network (CNN) on the downscaled input and a second, lighter CNN on the original input, merging their outputs to create a multiscale representation without significant computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple CNN computations are performed at different scales to create multiscale representation, then object recognition accuracy is improved, but computational cost increases prohibitively
Solution Approach 1:
The patent segments the computational task by running a single deep CNN at multiple scaled-down resolutions rather than running multiple complete CNNs at different scales. This divides the computational workload into one heavy computation performed multiple times on progressively smaller inputs, achieving multiscale representation with reduced total computational cost.
Solution Approach 2:
The patent introduces a scaling dimension by processing the input image at multiple resolution levels (e.g., original size, half size, quarter size) and combining the features from these different dimensional representations. This allows the system to capture objects at various scales without computing separate full-resolution CNNs for each scale.
2Productivity
If intermediate-level features are combined with deep-layer features to form richer representation, then computational efficiency is improved, but feature discriminative power decreases
Solution Approach 1:
The patent applies local quality by using different CNN depth levels for different spatial regions and object scales. Shallow layers capture local detailed features for small objects, while deep layers capture global semantic features for large objects. This localized application of different network depths optimizes both efficiency and discriminative power.
Solution Approach 2:
The patent dynamically adjusts the effective network depth for different input scales through the scaling operation. As the input image is downsampled, the same physical CNN layers effectively become shallower relative to the reduced input size, allowing the system to adaptively use appropriate feature depths for each scale without manual configuration.
Data Source
AI summary
Embodiments of the present invention are directed to a computer-implemented method for multiscale representation of input data. A non-limiting example of the computer-implemented method includes a processor receiving an original input. The processor downsamples the original input into a downscaled input. The processor runs a first convolutional neural network (“CNN”) on the downscaled input. The processor runs a second CNN on the original input, where the second CNN has fewer layers than the first CNN. The processor merges the output of the first CNN with the output of the second CNN and provides a result following the merging of the outputs.