ISO-DeTr: A novel detection transformer for industrial small object detection

Effectively detecting and assessing real-time structural and ecological parameters in contemporary manufacturing environments poses significant challenges, particularly in identifying minute objects within product images. The swift evolution of the industrial sector underscores the necessity for intelligent manufacturing environments to uphold stringent product quality standards. However, accelerating production processes at high speeds heightens the risk of defective product outcomes. This research addresses the challenges inherent in small object detection within industrial contexts, proposing an innovative detection transformer model tailored to modern manufacturing environments. The proposed model integrates a feature-enhanced multi-head self-attention block (FEMSA), merging cross-channel communication network and multiple multi-head self-attention (MSA) components to refine image features. A query proposal network is also introduced within the detection transformer framework to discern high-ranking proposals using Intersection over Union (IoU) and Non-Maximum Suppression (NMS) algorithms. Through extensive experimentation on custom industrial small objects, our proposed model demonstrates superior performance compared to existing models based on Non-Maximum Suppression and transformers. By tackling the challenges associated with small object detection, our model contributes to the dynamic synchronization between virtual and physical manufacturing realms, enhancing quality control in industrial production. • A novel detection transformer optimizes small object detection for industrial use. • Features C3Net integration for enhanced communication and reduced computation. • Introduces IoU-based assignments to replace Hungarian matching loss. • Simplifies training with a robust, end-to-end architecture using IoU loss. • Achieves higher precision and faster convergence on industrial datasets.