Multi-Level Feature Fusion for Continual Learning in Visual Quality Inspection

TL;DR

The paper tackles continual learning for visual quality inspection in volatile remanufacturing environments, where product variants and defect patterns shift. It introduces multi-level feature fusion (MLFF), a method that freezes a pretrained backbone and fuses representations from multiple depths through lightweight adapters and a small MLP to maintain performance with far fewer trainable parameters. Empirical results show MLFF matches or exceeds end-to-end training across several datasets while reducing forgetting in rehearsal-based continual learning and improving generalization to new product types and defects. The approach offers a practical, compute-efficient solution for edge deployment in manufacturing quality control.

Abstract

Deep neural networks show great potential for automating various visual quality inspection tasks in manufacturing. However, their applicability is limited in more volatile scenarios, such as remanufacturing, where the inspected products and defect patterns often change. In such settings, deployed models require frequent adaptation to novel conditions, effectively posing a continual learning problem. To enable quick adaptation, the necessary training processes must be computationally efficient while still avoiding effects like catastrophic forgetting. This work presents a multi-level feature fusion (MLFF) approach that aims to improve both aspects simultaneously by utilizing representations from different depths of a pretrained network. We show that our approach is able to match the performance of end-to-end training for different quality inspection problems while using significantly less trainable parameters. Furthermore, it reduces catastrophic forgetting and improves generalization robustness to new product types or defects.

Figures (5)

• Figure 1: Overview of the proposed approach. As an input image $x$ is processed by the pretrained feature extractor, intermediate representations are extracted at different stages and processed by the MLFF module.
• Figure 2: F1-Scores obtained on the test data of Re-GBC, SDNET, and NEU-DET using the MLFF approach, fully trained models, and finetuning of the linear and MLP classifiers.
• Figure 3: AF1 after continual learning on the Re-GBC dataset for different model architectures and numbers of historic samples.
• Figure 4: AF1 (+/- one standard deviation) obtained with a ResNet50 and the MLFF approach on the Re-GBC dataset for different rehearsal methods.
• Figure 5: t-SNE embeddings of the new, historic, and selected historic samples during the last adaptation round for different rehearsal methods using a ResNet50 and the MLFF approach (best viewed in color).