TL;DR
Hyundai Motor Company has developed an AI-powered application for verifying wheel design similarity, significantly reducing the design process time by 90%. This system leverages deep metric learning with a novel application of cross-entropy loss on a custom-built dataset of global wheel designs, demonstrating a practical solution for a real-world industrial challenge.
What Problem Does This Paper Address?
The automotive industry faces intense pressure to innovate and shorten design cycles due to rapidly changing consumer preferences. A critical and time-consuming part of this process is ensuring that new wheel designs do not infringe on existing patents or resemble competitor designs too closely. Traditionally, this verification was a manual, labor-intensive task relying on designers' experience, internet searches, and attending motor shows. This manual approach is not only slow but also prone to errors, potentially leading to costly legal issues and the mass production of infringing designs, with sunk costs exceeding $3,000,000. Hyundai Motor sought to automate and accelerate this crucial verification step.
Methods
The core of the solution is a deep metric learning (DML) model trained to extract feature vectors from wheel images. The application utilizes a ResNet50 backbone for feature extraction. A key methodological contribution is the use of a standard cross-entropy loss for DML, a departure from typical pairwise loss functions. The authors argue that cross-entropy loss, while traditionally used for classification, can achieve similar results to pairwise losses in DML by minimizing the conditional entropy H(Z|Y) and maximizing mutual information, effectively enforcing similar distances between samples in the embedding space.
The process involves:
- Image Feature Extraction: An AI model (ResNet50 trained with DML) extracts a feature vector from a query wheel image.
- Database Search: This feature vector is then used to search a large database of wheel images for similar feature vectors using an approximate search method (k-nearest neighbor graph).
- Similarity Verification: The application returns the top 6 most similar wheel images, along with their brand model and a similarity score, within seconds.
Datasets
- Custom Wheel Image Database: The researchers constructed a novel dataset by collecting approximately 730,000 automobile images from a large online image website. From these, they extracted wheel images using the Mask R-CNN algorithm. After filtering for quality, they compiled a database of over 500,000 wheel images, representing global brands.
- Labeling: Instead of traditional feature-based labels, they used the car's brand model as the label for each wheel image. This was a practical solution to the challenge of manually labeling such a large dataset.
- Transfer Learning: A Semi-weakly Supervised Learning (SWSL) ImageNet model from Facebook AI Research (FAIR) was used for transfer learning.
Key Insights and Ideas
- Cross-Entropy Loss for DML: The paper empirically demonstrates that cross-entropy loss, often overlooked for DML, can be as effective as pairwise losses in learning discriminative embeddings. This opens up new avenues for loss function selection in DML tasks.
- Brand Model as Label: Using the car's brand model as a label for wheel images is an innovative and practical approach to dataset labeling for DML, especially when dealing with large, real-world datasets where fine-grained feature labels are scarce.
- Practical AI Deployment: The successful integration of the DML model into a user-friendly web service using AWS highlights the feasibility of deploying complex AI solutions for immediate industrial impact.
- Edge Detection for Sketched Designs: The inclusion of edge detection in the image processing pipeline allows the system to effectively verify similarity for both rendered wheel images and hand-sketched designs, broadening its applicability.
Key Results
- Time Reduction: The application reduced the wheel design similarity verification time by 90%, from approximately 120 minutes per design to a maximum of 10 minutes.
- Recall@1 Performance: The DML model achieved a Recall@1 score of around 88.93% at epoch 6000, indicating high accuracy in identifying similar wheel designs.
- Database Size: The constructed database contains over 500,000 wheel images from global markets.
- Speed: The verification process (searching and returning results) takes between one to three seconds.
Limitations and Future Work
- Model Performance vs. Epoch: The paper notes that at lower epochs, the model focuses more on design details (like color), while at higher epochs, it prioritizes overall shape. The chosen epoch (6000) represents a balance between these two aspects.
- Dataset Size: While 500,000 images is substantial, the paper doesn't explicitly discuss the potential for further improvements with even larger or more diverse datasets.
- Future Work: The authors suggest that this approach can be extended to other design verification tasks within the automotive industry. The feedback mechanism using a "dislike" button in the deployed application (Figure 5) indicates a clear path for continuous model improvement through active learning.
Summary updated