On-Device Computer Vision Filtering for Lightweight Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

State-of-the-art object detectors are computationally expensive and complex, making them challenging for use on local devices like smartphones, and existing methods for improving precision and recall in object detection often increase complexity further.

Innovation Solution

Implement a two-stage approach using a lightweight initial filter to classify image regions, sending only regions of interest to a more specialized model, and masking or cropping regions without objects to reduce computational load.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning models are trained to recognize objects, actions, and scenes, then computer vision performance is improved, but model size and computational requirements increase

Engineering Contradiction:
Improvecomputer vision performanceVSAvoidmodel size
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the machine learning model into a teacher model (large, high-accuracy model used for training) and a student model (small, efficient model deployed on-device). The teacher model processes images to generate training data, which is then used to train the compact student model that runs on mobile devices with limited resources.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates a simplified copy of the large teacher model by training a student model on distilled knowledge from the teacher. The student model replicates the essential vision recognition capabilities while occupying significantly less memory and requiring fewer computational resources for inference on mobile devices.

Inventive Principle:
Principle #26Copying

2Measurement precision

If large machine learning models are deployed on mobile devices, then vision recognition accuracy is improved, but memory usage and processing power requirements increase

Engineering Contradiction:
Improvevision recognition accuracyVSAvoidprocessing power requirements
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system separates the computationally intensive training phase (performed on server using large teacher model) from the inference phase (performed on mobile device using small student model). This segmentation allows high accuracy to be achieved during training while keeping on-device processing requirements minimal.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The student model is self-sufficient once trained, requiring minimal processing power to run on the mobile device. The model performs object, action, and scene recognition independently without needing continuous connection to external computing resources, enabling efficient on-device operation.

Inventive Principle:
Principle #25Self-service

3Difficulty of detecting and measuring

If comprehensive object recognition is implemented, then detection capability is improved, but on-device storage requirements increase

Engineering Contradiction:
Improvedetection capabilityVSAvoidon-device storage
Core Design Contradiction:
Difficulty of detecting and measuringVSQuantity of substance

Solution Approach 1:

The patent creates a compressed knowledge copy from the large teacher model that fits within mobile device storage constraints. The student model captures the essential detection capabilities for objects, actions, and scenes in a compact form factor, enabling comprehensive recognition without requiring gigabytes of additional storage space.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4091089B1Systems and methods for improved computer vision in on-device applications
Publication Date: 2026.05.13 GOOGLE LLC
  • EP4091089B1 patent drawingFigure 1a
  • EP4091089B1 patent drawingFigure 1B
  • EP4091089B1 patent drawingFigure 1C

AI summary

The present disclosure is directed to object and/or character recognition for use in applications such as computer vision. Advantages of the present disclosure include lightweight functionality that can be used on devices such as smart phones. Aspects of the present disclosure include a sequential architecture where a lightweight machine-learned model can receive an image, detect whether an object is present in one or more regions of the image, and generate an output based on the detection. This output can be applied as a filter to remove image data that can be neglected for more memory intensive machine-learned models applied downstream.