Introduction
Fine-Grained Visual Classification (FGVC) represents a significant challenge in computer vision, requiring the differentiation of highly similar sub-categories. In industrial contexts, such as manufacturing and anti-counterfeiting, data typically arrives in sequential batches over time. This temporal progression creates a strategic dilemma: how to best integrate new data with existing models without losing performance or wasting computational resources. The Dual-Carriageway Framework (DCF) addresses this by automating the selection of optimal training strategies and image preprocessing techniques, specifically focusing on the often-overlooked impact of image padding.
Methodology: The Dual-Carriageway Framework
The DCF is designed to systematically evaluate the best "pathway" for model evolution as new datasets become available. It operates in two primary stages: the Padding Scheme Adaptor (PSA) and the Training Pathway Selector (TPS).
Padding Scheme Adaptor (PSA)
Standard preprocessing often involves padding non-square images to fit model input requirements. However, in FGVC, edge artifacts can interfere with the detection of subtle patterns. The PSA evaluates six padding methods: constant values (zero, white, and grey), mean-based values (RGB-mean and LAB-mean), and reflection padding, which mirrors the image content at the borders to maintain feature continuity.
Training Pathway Selector (TPS)
Once the optimal padding is determined, the TPS evaluates five distinct training settings to find the most effective way to incorporate new data (Dataset B) with old data (Dataset A):
- Setting 1: Training exclusively on Dataset A.
- Setting 2: Sequential training (pre-training on A, then fine-tuning on B).
- Setting 3: Joint training (simultaneous training on A and B).
- Setting 4: Training exclusively on Dataset B.
- Setting 5: Reverse sequential training (pre-training on B, then fine-tuning on A).
Experimental Setup
The framework was validated using real-world industrial datasets provided by Procter & Gamble, consisting of product bottom images for manufacturer identification ('F1' vs 'F2').
- Dataset A: 5,870 images representing the initial data state.
- Dataset B: 3,112 images collected later, introducing variations in lighting, noise, and subtle pattern shifts.
The evaluation utilized popular architectures including ResNet18, ResNet34, and Inception-v3 to ensure model-agnostic results.
Results and Discussion
The experimental results highlighted two critical factors for success in sequential FGVC tasks: the choice of padding and the sequence of learning.
The Impact of Padding
Reflection padding emerged as the superior method, consistently outperforming standard zero-padding. By avoiding artificial edges, reflection padding allows the model to focus more effectively on the actual regions of interest.
Quantitatively, reflection padding improved testing accuracy by an average of 3.72%. Visual explanations using GradCAM confirmed that this method helps the model maintain focus on the sparse, fine-grained code patterns rather than being distracted by the "dead space" of constant-value padding.
Pathway Optimization
The Training Pathway Selector revealed that the order of learning is vital. Setting 2 (fine-tuning from the older, larger Dataset A to the newer Dataset B) generally provided the most robust performance. This suggests that establishing a broad feature base on initial data before specializing on newer nuances is more effective than training from scratch on new data alone (Setting 4), which it outperformed by up to 2.13%.
Conclusion
The Dual-Carriageway Framework provides a systematic, explainable approach to evolving FGVC models in industrial settings. By treating padding as a critical hyperparameter and evaluating multiple training pathways, DCF ensures high accuracy and provides insights into model behavior. Future work will focus on reducing the computational overhead of testing multiple pathways and extending the framework to handle continuous streams of data in a lifelong learning context.
Summary updated