CNN Activation Segmentation for Object Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine vision systems for object classification rely on studying neural activations in single units, require human feedback, and neglect neural activation patterns over the entire network, leading to inefficiencies in automatic inspection and classification.

Innovation Solution

A system that identifies salient regions in an intensity image using convolutional neural network (CNN) activation patterns, generates multi-scale features through general average pooling, and employs unsupervised clustering to map key attributes onto probabilities for object categorization, enabling automated object classification and control of vehicle components.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If neural activations in single units are studied for object classification, then the system can process individual neuron responses, but it fails to capture neural activation patterns over the entire network and requires human feedback

Engineering Contradiction:
Improveobject classification accuracyVSAvoidautomation level
Core Design Contradiction:
Measurement precisionVSExtent of automation

Solution Approach 1:

The patent segments the CNN into multiple layers and analyzes activation patterns at each layer separately. By dividing the network into discrete layers and examining activation patterns at each level, the system captures hierarchical feature representations without requiring human feedback, thus improving both classification accuracy and automation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent transitions from analyzing single-unit activations to examining activation patterns across the entire network by introducing a spatial dimension (layer position) to the analysis. This dimensional expansion allows the system to capture distributed activation patterns across multiple layers, enabling automated classification without human intervention.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Device complexity

If single-unit neural activations are analyzed for object classification, then the processing focus is simplified, but the system neglects neural activation patterns over the entire network

Engineering Contradiction:
Improveanalysis complexityVSAvoidneural activation pattern information
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent merges activation patterns from multiple CNN layers by concatenating feature vectors extracted at different layer depths. This combination integrates information from early feature detection layers with higher-level semantic layers, preserving comprehensive neural activation patterns while maintaining manageable analysis complexity through systematic feature fusion.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent creates a universal feature representation that serves multiple classification purposes by aggregating activation patterns across all network layers. This multi-functional feature set can be used for various object classification tasks while capturing the full spectrum of neural activation information, eliminating information loss.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If human feedback is required to identify common themes in top scoring regions, then classification accuracy can be improved, but productivity and automation are reduced

Engineering Contradiction:
Improveclassification accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent implements self-service classification by using the CNN's own activation patterns across multiple layers to automatically identify and classify objects. The system leverages the network's inherent feature extraction capabilities at different depths to perform classification without external human feedback, thereby maintaining high accuracy while dramatically improving processing speed and productivity.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent introduces internal feedback loops where activation patterns from deeper layers are used to refine and adjust the interpretation of features from earlier layers. This self-feedback mechanism enables the system to automatically converge on accurate classifications without human intervention, simultaneously improving both accuracy and productivity.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3602489B1Machine-vision method to classify input data based on object components
Publication Date: 2023.08.23 HRL LAB

AI summary

Described is a system for classifying objects and scenes in images. The system identifies salient regions of an image based on activation patterns of a convolutional neural network (CNN). Multi-scale features for the salient regions are generated by probing the activation patterns of the CNN at different layers. Using an unsupervised clustering technique, the multi-scale features are clustered to identify key attributes captured by the CNN. The system maps from a histogram of the key attributes onto probabilities for a set of object categories. Using the probabilities, an object or scene in the image is classified as belonging to an object category, and a vehicle component is controlled based on the object category causing the vehicle component to perform an automated action.