An enhanced strategy for query-based defect detector via adaptive spatial feature Reorganization And cross-stage query Injection

Abstract

Detection of surface defects from images is crucial to ensure high quality products in manufacturing applications, where surface detection of small defects plays a vital role and has received much attention in the manufacturing industry. However, existing detection solutions perform unevenly in different small defect scenarios. Therefore, this paper proposes an efficient enhancement strategy (RAI) to enhance the model’s ability to detect small surface defects. It consists of two major parts: (i) the feature information enhancement part (ASFR), which consists of a frequency balance (FB) module, an adaptive dilation convolution kernel (ADCK) module, and a spatial feature reorganization (SFR) module, to progressively enhance the semantic information of small defects; and (ii) the subsequent-stage correction interpretation part, which consists of a cross-stage query injection (CQI) mechanism to correct the training focus imbalances and the cascading errors, and fine-grained interpretation of minor defect features. On the engineering side, we applied the strategy to Deformable-Detection Transformer (DETR), Dynamic Anchor Boxes-DETR, and Adamixer, based on three datasets: a self-constructed bamboo slice defect dataset, a defect dataset from Northeastern University, and huggingface surface defects. The experiments were conducted, and mAP50 was improved by 1.7% to 12.5% on the bamboo slice defect test set, 2.4% to 7.8% on the NEU-DET test set, and 2.8% to 4.0% on the huggingface surface defect test set.