DCEI-RTDETR: an improved RT-DETR-based detection algorithm for data center equipment indicator lights

• Published in: Journal of real-time image processing Vol. 22; no. 1; p. 24
• Main Authors: Xu, Liangjie, Zeng, Wenxian
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.02.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Computer centers; Computer Graphics; Computer Science; Control equipment; Data centers; Datasets; Deep learning; Design; Feature extraction; Feature maps; Image Processing and Computer Vision; Localization; Modules; Multimedia Information Systems; Occlusion; Optimization; Pattern Recognition; Query processing; Scaling factors; Semantics; Signal,Image and Speech Processing; Target detection; Training; Inner-EIOU; High-level screening feature aggregation network; One-to-many label assignment; DCEI-RTDETR
• ISSN: 1861-8200; 1861-8219
• DOI: 10.1007/s11554-024-01599-2

To address the issue of low detection accuracy caused by small target sizes and complex image backgrounds in the detection of indicator lights in data center equipment rooms, a lightweight small-target detection algorithm based on the Transformer architecture, named DCEI-RTDETR, is proposed. First, EfficientFormerV2 is utilized as the feature extraction network. By reducing the number of downsampling operations, the size of the feature maps is increased, enabling the network to focus more effectively on small targets. Subsequently, a high-level screening feature aggregation network is designed, which employs the HiLo attention mechanism at the highest feature level as an intra-feature interaction module. Simultaneously, the GSConv and VoVGCSCSP modules are used for cross-scale feature fusion, dynamically optimizing the feature map’s representation capability. Additionally, a one-to-many label assignment method is introduced, utilizing a grouped decoder to optimize object query processing, thereby alleviating the issues of occlusion and the loss of small target feature information. Finally, the GIOU loss function is replaced with Inner-EIOU incorporating a scaling factor to control the auxiliary bounding box, which improves the accuracy of small target detection. Experimental results on a proprietary data center equipment status detection dataset show that, compared to the original RT-DETR algorithm, the mAP50 is increased by 4.2%, the mAP50:95 by 2.1%, and the FPS is maintained at 90. Generalization experiments on the public VisDrone2021 dataset also demonstrate the effectiveness and generality of the proposed algorithm, with the mAP50 improved by 4.4%.