Object detection for remote sensing is a fundamental task in image processing of remote sensing; as one of the core components, small or tiny object detection plays an important role. Despite the considerable advancements achieved in small object detection with the integration of CNN and transformer networks, there remains untapped potential for enhancing the extraction and utilization of information associated with small objects. Particularly within transformer structures, this potential arises from the disregard of the complex and the intertwined interplay between spatial context information and channel information during the global modeling of pixel-level information within small objects. As a result, valuable information is prone to being obfuscated and annihilated. To mitigate this limitation, we propose an innovative framework, YOLO-DCTI, that capitalizes on the Contextual Transformer (CoT) framework for the detection of small or tiny objects. Specifically, within CoT, we seamlessly incorporate global residuals and local fusion mechanisms throughout the entire input-to-output pipeline. This integration facilitates a profound investigation into the network’s intrinsic representations at deeper levels and fosters the fusion of spatial contextual attributes with channel characteristics. Moreover, we propose an improved decoupled contextual transformer detection head structure, denoted as DCTI, to effectively resolve the feature conflicts that ensue from the concurrent classification and regression tasks. The experimental results on the Dota, VISDrone, and NWPU VHR-10 datasets show that, on the powerful real-time detection network YOLOv7, the speed and accuracy of tiny targets are better balanced.

Calculating semantic similarity is paramount in medical information processing, and it aims to assess the similarity of medical professional terminologies within medical databases. Natural language models based on Bidirectional Encoder Representations from Transformers(BERT) offer a novel approach to semantic representation for semantic similarity calculations. However, due to the specificity of medical terminologies, these models often struggle with accurately representing semantically similar medical terms, leading to inaccuracies in term representation and consequently affecting the accuracy of similarity calculations. To address this challenge, this study employs Chat Generative Pre-trained Transformer (ChatGPT) and contrastive loss during the training phase to adapt BERT, enhancing its semantic representation capabilities and improving the accuracy of similarity calculations. Specifically, we leverage ChatGPT-3.5 to generate semantically similar texts for medical professional terminologies, incorporating them as pseudo-labels into the model training process. Subsequently, contrastive loss is utilized to minimize the distance between relevant samples and maximize the distance between irrelevant samples, thereby enhancing the performance of medical similarity models, especially with limited training samples. Experimental validation is conducted on the open Electronic Health Record (OpenEHR) dataset, randomly divided into four groups to verify the effectiveness of the proposed methodology.

Open-world object detection aims to identify unseen instances as unknown and support the continuous learning of newly introduced categories. However, in remote sensing imagery, severe foreground-background ambiguity complicates the discovery of unknown objects and exacerbates catastrophic forgetting during incremental learning. To this end, we propose CTKR-DETR, a Transformer-based detector that jointly enhances unknown-object discovery and historical-knowledge retention. First, we leverage the Segment Anything Model with an objectness-aware one-to-many label assignment to filter background noise and establish robust supervision for potential unknown objects. Second, a quality-guided IoU-aware classification loss provides continuous localization-quality supervision for known objects and applies a rank-based dynamic decay for unknown targets, decoupling their optimization. Finally, a cross-task knowledge rehearsal strategy uses a teacher model to consolidate prior object representations. Extensive experiments on DOTA 1.0, DOTA 1.5, and ShipRSImageNet demonstrate CTKR-DETR achieves superior performance in unknown-object recall and known-object detection.