Transformer Based Deep Learning Model For Object Detection Using Meta Learning

Object detection is the identification and pinpointing of the exact location of objects that belong to particular categories inside an image. However, insufficient annotated data in real-world applications hinders the training of accurate deep-learning models for object detection. To address this problem, models are trained using meta-learning based few-shot learning approach specifically designed for object recognition. Although these procedures are effective, yet have limitations due of inadequate geographic proposals for new classes and insufficient evaluation of interclass linkages. These limitations restrict the transfer of information from existing to novel classes. The Meta-DETR framework incorporates correlational aggregation into DETR frameworks, improving object detection performance in few-shot scenarios. We propose a modified Meta-DETR model that substitutes the initial aggregation module of RESNET152 with a tailored variant. This modification efficiently captures the interconnections between several classes, minimizing misclassifications and enhancing knowledge transfer to novel classes. The experimental results confirm that our Meta-DETR model outperforms state-of-the-art methods on different few-shot object detection benchmarks.