This letter focuses on the fact that small objects with few pixels disappear in feature maps with large receptive fields, as the network deepens, in object detection tasks. Therefore, the detection of dense small objects is challenging. A DI-YOLOv5 object detection algorithm is proposed. Specifically, a dual-wavelet convolution module (DWCM), which contains DWT_Conv and IWT_Conv, is proposed to reduce the loss of feature map information while obtaining feature maps with a large receptive field. The DWT _ Conv and IWT _ Conv can be used as replacements for downsampling and upsampling operations. Moreover, in the process of information transmission to the deep layer, a CSPCoA module is proposed to further capture the location information and information dependencies in different spatial directions. DWCM and CSPCoA are single, generic, plug-and-play units. We propose DI-YOLOv5 with YOLOv5 [1] as the baseline, and extensively evaluate the performance of these two modules on small object detection. Experiments demonstrate that DI-YOLOv5 can effectively improve the accuracy of object detection.

The existing research on open-set recognition (OSR) and novel class discovery (NCD) has largely treated these tasks as independent fields. OSR aims to identify samples that do not belong to the training set classes, while NCD seeks to further classify such unseen, unlabeled samples into novel classes. However, there is a lack of a unified framework to automate both tasks systematically. In this paper, we propose a unified training framework to identify and categorize unseen samples. Specifically, we conduct a comprehensive evaluation of existing post hoc OSR methods and observe that their performance is highly sensitive to the temperature scaling factor. To address this, we introduce a distance-based evaluation method for OSR, which not only outperforms existing post hoc approaches but also integrates seamlessly with them to deliver enhanced performance. Furthermore, we developed a prototype-based classification head leveraging this distance metric, which facilitates compact feature representations for known classes and guides the clustering of unknown classes, thereby significantly enhancing the classification accuracy for novel classes. On the CUB-200-2011 dataset, our unified framework achieves a 0.95–6.12% improvement in AUROC scores on OSR benchmarks and a 3.19% increase in classification accuracy for novel classes. Extensive experiments and visualizations validate the effectiveness of the proposed approach. We believe that this unified framework will pave the way for automating the integration of OSR and NCD, offering a more efficient and systematic approach to addressing these tasks.

Mobile vision applications require backbones that balance accuracy, efficiency, and model complexity under strict resource constraints. We propose GAMLA, a lightweight backbone that integrates local feature enhancement with efficient global dependency modeling. The architecture introduces a Gated Aggregation (GA) block to fuse multi-scale features and reduce channel redundancy, and a Modulated Linear Attention (MLA) block to capture long-range dependencies with minimal overhead. Extensive experiments across image classification, object detection, instance segmentation, and semantic segmentation demonstrate that GAMLA achieves an optimal trade-off between accuracy and efficiency. In early model scaling (from tiny to small), GAMLA improves scaling parameter efficiency (SPE) by approximately 93% and scaling cost-effectiveness (Score_coeff) by around 81% over the state-of-the-art iFormer. Medium-scale GAMLA models exhibit strong and stable performance on downstream tasks under tight resource constraints, making it ideal for mobile and edge vision applications requiring both efficiency and accuracy.