Surface Defect Detection Using Defect Score Matrices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based methods for detecting surface defects in cold-rolled steel sheets face challenges due to diverse defect characteristics and image sizes, leading to feature loss and reduced classification reliability, particularly when dealing with defects of varying shapes and types.
Innovation Solution
A method involving the generation of defect matrices and scores to accurately detect and classify defects by calculating first and second defect scores, applying standard deviation, and generating defect images using reference windows, while adjusting image sizes through padding or removal to maintain feature integrity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional preprocessing methods (crop, resize, padding) are used to standardize defect images, then image size is standardized, but image features are lost and classification reliability is degraded
Solution Approach 1:
The patent changes the parameter representation from raw pixel values to defect scores derived from multiple detection models. Instead of resizing images (which loses features), it transforms the image data into a defect score matrix where each element represents the likelihood of a defect at that position, calculated by aggregating predictions from multiple SSD models with different architectures.
Solution Approach 2:
The patent introduces defect score matrices as an intermediary representation between the original defect images and the classification model. These matrices serve as a feature-rich intermediate form that preserves spatial defect information while being standardized in size, allowing the classification model to work with uniform inputs without losing critical defect characteristics.
2Reliability
If multiple defect detection models are used to improve detection accuracy, then detection reliability is improved, but processing time and computational complexity increase
Solution Approach 1:
The patent performs preliminary defect detection using multiple SSD models before the classification stage. By pre-calculating defect score matrices from multiple detection models and aggregating their predictions, the system prepares comprehensive defect information in advance, allowing the subsequent classification model to make accurate decisions more efficiently without re-processing raw images multiple times.
3Reliability
If defect images with various sizes and aspect ratios are processed directly, then image integrity is maintained, but deep learning model processing is hindered
Solution Approach 1:
The patent transitions from processing images in the spatial dimension (varying sizes and aspect ratios) to processing defect information in a standardized score matrix dimension. The defect score matrices are all of the same size regardless of the original image dimensions, allowing uniform processing by deep learning models while preserving defect information through the score representation rather than pixel values.
Data Source
Figure 1
Figure 2
Figure 3(a)~3(g)
AI summary
According to an embodiment of the present application, there is provided a method of detecting a surface defect. The method includes acquiring a target image, detecting at least one defect area from the target image, generating a first defect matrix having a size corresponding to the target image and an element that is a first defect score calculated based on the number of defect areas, calculating a plurality of second defect scores by summing first defect scores in a range corresponding to a predetermined reference window in the first defect matrix, and extracting an image corresponding to the reference window from the target image based on the second defect score and generating a defect image.