Ψ-Net is an Efficient Tiny Defect Detector

Abstract

Detecting tiny defects is not easy for industrial product manufacturers. Classical models have witnessed remarkable progress in detecting defects. Nevertheless, these models may fail to detect tiny defects effectively and efficiently when trained with few samples. Accordingly, we propose an efficient two-stage detector, Ψ-Net, to solve the problem of detecting tiny defects with few samples. In the first stage, we present an efficient model for extracting region proposals from large images. In the second stage, we propose the ResNeLt integrated with Ψ-Attention to classify the region proposals. ResNeLt is a lightweight network that enables the model to be trained with few samples. Meanwhile, Ψ-Attention, a model-agnostic plug-in, improves the featureencoding capability of tiny defects detectors. The proposed model outperforms several state-of-the-art models on NEU-CLS dataset. In addition, the accuracy and sensitivity of Ψ-Net have been significantly improved with few samples over Surface Crack Detection dataset and our own-collected Robber-S dataset.