Unified detection method of aluminium profile surface defects: Common and rare defect categories

Abstract It is difficult to achieve automatic visual detection of aluminium profile surface defects (APSD) owing to their various categories, irregular shapes, random distribution, and unbalanced samples. Utilising the attention mechanism, the unified detection method attempts to address these challenges for both common and rare defects. We formulate our method as a variant of few-shot learning to recognise the common and rare defect categories. First, a category representation network is applied to extract common category feature maps (CCMs). Second, an attention module is proposed to generate the proposal feature maps (PMs) of each rare category. Third, rare category feature maps (RCMs) are transformed from the CCMs under the guidance of the PMs. Finally, the scores of each category are obtained through the spatial pooling of both CCMs and RCMs. Experimental results on our constructed dataset show that our method is effective and outperforms the state-of-the-art methods.