Knowledge Distillation for Object Detection in Blurry Images

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing object detection techniques struggle with blurry images due to noise and motion blur, leading to degraded performance in commercially available camera images, as datasets like PACAL VOC and COCO do not adequately consider these issues.

Innovation Solution

A knowledge distillation-based system is employed, where a student network is trained using a teacher network to improve detection performance in blurry images by incorporating a deblurring subnet with 1×1 and 3×3 deformable convolution kernels, and a total loss function that includes deblurring and self-guided learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional object detection techniques are used on blurry images, then the system structure remains simple, but detection accuracy deteriorates due to noise and motion blur

Engineering Contradiction:

Improvedetection accuracyVSAvoidsystem complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The detection system is divided into two specialized networks: a teacher network optimized for sharp images and a student network optimized for blurry images. Each network has its own feature extraction and detection subnets, allowing them to specialize in different image conditions without requiring a complex unified architecture.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

A knowledge distillation mechanism serves as an intermediary between the teacher and student networks. The teacher network provides soft labels and gradient information to guide the student network's training, enabling the student to learn from the teacher's expertise while maintaining independence for blurry image processing.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Reliability

If a single detection network is used for both sharp and blurry images, then the system structure remains simple, but detection performance on blurry images deteriorates

Engineering Contradiction:

Improvedetection performance on blurry imagesVSAvoidnetwork architecture complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The student network incorporates deformable convolution kernels with learnable offset fields that can dynamically adjust kernel positions and shapes. This allows the network to adapt its parameters to handle motion blur and noise specifically, improving reliability on blurry images while maintaining a relatively simple overall structure.

Inventive Principle:

Principle #35Parameter changes

3Measurement precision

If knowledge distillation is applied to train student network from teacher network, then detection accuracy on blurry images improves, but training time and computational resources increase

Engineering Contradiction:

Improvedetection accuracyVSAvoidtraining time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The knowledge distillation process focuses on transferring only the most critical knowledge from teacher to student network, using soft labels and gradient information rather than full model replication. This partial action approach achieves sufficient accuracy improvement without requiring excessive training time or computational resources.

Inventive Principle:

Principle #16Partial or excessive action

Solution Approach 2:

The knowledge distillation is implemented as a continuous training process where the student network progressively learns from the teacher network throughout training. This continuous transfer of knowledge allows the student to gradually improve detection accuracy on blurry images without requiring separate lengthy training phases, reducing overall training time.

Inventive Principle:

Principle #20Continuity of useful action

Data Source

PatentUS12530749B2Knowledge distillation-based system for learning of teacher model and student model

Publication Date: 2026.01.20 KOREA UNIV RES & BUSINESS FOUND

AI summary

Disclosed is a system for improving the detection performance of objects of interest in blurry images. According to an embodiment of the disclosed system for improving the detection performance of objects of interest in blurry images, the system includes a teacher network including a first base subnet, a first feature pyramid subnet and a first detection subnet; and a student network including a second base subnet, a second feature pyramid subnet, a deblurring subnet and a second detection subnet, wherein the teacher network provides hints to the student network.