Abstract:Visual surface defect inspection provides an important tool for product quality assessment in a wide range of industrial applications. In recent years, diverse convolutional neural network (CNN) models have been devel-oped for the high-accuracy image classification of natural objects. However, it remains a challenging task to develop deep-learning-based approaches for surface defect inspection of industrial products. The main challenge is that surface defects of industrial products occur in low probability, therefore it is impractical to build a large image dataset containing different types of defects to train/fine-tune deep CNN models to achieve satisfactory performance when applied to industrial inspection tasks. To overcome the above-mentioned problem, we build a light-weight defect classification model based on the pre-trained SqueezeNet architecture and present three effective techniques to achieve high-accuracy defect classification using very few collected defective training samples. First, we experimentally demonstrate that it is reasonable to freeze shallower convolutional layers and only fine-tune deeper layers in the last convolutional stage when there are not enough training samples. Second, we integrate batch-size independent Group Normalization (GN) into SqueezeNet to compute stable statistics based on a limited number of defective samples for training the classification model. Third, we add an auxiliary task to perform the similarity comparison of two defective samples during the training phase, significantly increasing the size of the training data and generating more distinctive defect-specific representations. The experimental results on the NEU-CLS dataset and USB-FS dataset verify the effectiveness of the proposed method. Our method can obtain accurate defect classification results when very few defective training samples are provided, i.e., about 97.69% and 82.92% average classification accuracy on the NEU-CLS and USB-FS datasets, respectively, under the 5-shot setting (5 images per defect category).

Optimized CNN Model for Identifying Similar 3D Wear Particles in Few Samples

Automated 3D Ferrograph Image Analysis for Similar Particle Identification with the Knowledge-Embedded Double-Cnn Model

Knowledge-Guided Convolutional Neural Network Model for Similar Three-Dimensional Wear Debris Identification with Small Number of Samples

Recognition of Fatigue and Severe Sliding Wear Particles Using a CNN Model with Multi-Scale Feature Extractor

Wear Monitoring of Helical Milling Tool Based on One-Dimensional Convolutional Neural Network

A Hybrid Convolutional Neural Network for Intelligent Wear Particle Classification

Integrated Model of BP Neural Network and CNN Algorithm for Automatic Wear Debris Classification

Particle Identification in Particle Tracking Velocimetry Using Two-Stage Neural Networks

A Calibration-Informed Deep Learning Model for Three-Dimensional Particle Reconstruction of Volumetric Particle Image Velocimetry

A Partial Hierarchical Model for Online Low-Resolution Wear Particle Images Classification.

Intelligent classification of online wear particle in lubricating oil using optical direct imaging method and convolutional neural network for rotating machinery

Semantic Segmentation of Ferrography Images for Automatic Wear Particle Analysis

CBAM-YOLOv5: A Promising Network Model for Wear Particle Recognition

Shape classification of wear particles by image boundary analysis using machine learning algorithms

A Prototype of On-Line Extraction and Three-Dimensional Characterisation of Wear Particle Features from Video Sequence

A Cylinder Block Dynamic Characteristics-Based Data Augmentation Method for Wear State Identification under Data Imbalance Condition

Tool Wear Condition Monitoring Method Based on Deep Learning with Force Signals

An Effective Industrial Defect Classification Method under the Few-Shot Setting Via Two-Stream Training

Tool wear estimation using a CNN-transformer model with semi-supervised learning

Wear Depth Estimation from Single 2-D Image Based on Shape from Shading and Convolutional Neural Network Hybrid Model for In-Situ Wear Assessment

Wear particles image enhancement using long short-term memory 3D recurrent reconstruction neural network (LSTM 3D-R2N2)