Online Reflected-Light Ferrograph Image Enhancement for Wear-Particle Clarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing online reflected light ferrograph images suffer from low contrast, color bias, and blurred contours due to light scattering and contamination, hindering accurate wear particle feature extraction in machinery condition monitoring.
Innovation Solution
A method involving the fusion of SqueezeNet-Unet and ResNeXt-CycleGAN networks to enhance multiple features in online reflected light ferrograph images, using a SqueezeNet-Unet-based wear particle position prediction network and a ResNeXt-CycleGAN image transformation network, optimized with a weighted loss function for accurate feature enhancement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If traditional image enhancement algorithms (gray-scale transformation, histogram equalization) are used, then image contrast is improved, but color bias remains and contour features are blurred
Solution Approach 1:
The patent merges multiple enhancement algorithms (gray-scale transformation, histogram equalization, Retinex, AWB, ACE, sharpening techniques) into a unified deep learning framework that processes all image features simultaneously, resolving the contradiction by combining the strengths of each approach while eliminating their individual weaknesses through integrated multi-task learning
Solution Approach 2:
The patent transforms the enhancement process from manual parameter adjustment to automatic parameter optimization by training deep neural networks to learn optimal enhancement parameters from data, enabling simultaneous optimization of contrast, color, and contour features without the trade-offs of traditional fixed-parameter algorithms
2Loss of information
If color enhancement algorithms (Retinex, AWB, ACE) are used, then image color is improved, but contrast and contour features are not enhanced
Solution Approach 1:
The patent combines color enhancement algorithms with contrast enhancement and contour sharpening techniques within a single deep learning model, enabling simultaneous improvement of all three features through multi-objective optimization rather than sequential processing
3Shape
If Sharpening techniques (Sobel, Canny, Prewitt) are used, then image contour characteristics are improved, but contrast and color characteristics are not enhanced
Solution Approach 1:
The patent integrates contour sharpening operations with color correction and contrast enhancement in a unified neural network architecture, allowing all three enhancement objectives to be achieved simultaneously through joint optimization of multiple loss functions
4Loss of information
If existing enhancement models with multiple parameters are used, then image features are enhanced, but the models have poor generalization capability and require extensive parameter adjustment
Solution Approach 1:
The patent implements self-service by enabling the enhancement model to automatically learn and adjust its parameters from training data without manual intervention, using supervised learning to optimize enhancement parameters specific to each image while maintaining generalization capability across different datasets
Data Source
AI summary
A method and system of enhancing online reflected light ferrograph images. The method includes: based on contour markers of wear particles in the online reflected light ferrograph image, performing concatenate fusion on the SqueezeNet-Unet-based wear particle position prediction network and the ResNeXt-CycleGAN image transformation network to construct an online reflected light ferrograph image enhancement model; determining loss function of the position prediction network; combining SSIM and L1 losses to optimize cycle-consistency loss function of the ResNeXt-CycleGAN image transformation network; designing overall loss function of the ferrograph image enhancement model by weighted fusion; and optimizing the ferrograph image enhancement model with the overall loss function as optimization object successively using a training sample set consisting of an original online reflected light ferrograph image and a traditional algorithm-enhanced online reflected light ferrograph image and a training sample set consisting of the original image and an offline reflected light ferrograph image.