FastRecon++: Enhancing Few-Shot Anomaly Detection for Electric Vehicle Manufacturing

Ensuring product quality through accurate and efficient anomaly detection is essential in electric vehicle (EV) manufacturing. This paper presents FastRecon++, an improved few-shot anomaly detection approach built upon the original FastRecon framework. While FastRecon performs well on anomalous samples, it suffers from a high false positive rate on normal images. To address this issue, we introduce four key enhancements: (i) replacing the Wide ResNet-50 backbone with ResNet152 to capture richer feature representations; (ii) applying feature normalization to improve robustness and consistency; (iii) removing the average pooling layer to preserve fine-grained spatial information; and (iv) incorporating a threshold-based post-processing step to reduce noise in segmentation maps. These modifications significantly reduce false positives and improve both segmentation accuracy and image-level classification. Our method achieves strong results on the MVTec AD dataset and an EV-related Triangle dataset, demonstrating its potential for reliable industrial inspection, particularly in EV parts manufacturing.