Lightweight Self-reconstruction-based Feature Enhancement Framework for Real-time Small Object Detection

Abstract

We propose a framework that integrates a lightweight self-reconstruction-based feature enhancement module into the real- time detector YOLOv26) for improved small object detection. The self-reconstruction-based approach enhances feature representations of small objects by reconstructing the input image from the feature maps and utilizing the difference map between the reconstructed and original images as spatial prior information for small object localization. However, the computational overhead introduced by reconstructing the image up to the original resolution makes it difficult to apply directly this approach to real-time detectors. The purpose of the reconstructed image is not the achievement of perfect image restoration but the identification of the locations of information loss of small objects through the use of the difference map. Based on this analysis, we design a lightweight reconstruction head that generates difference maps at a lower resolution by reducing the number of upsampling operations and integrate it into YOLOv26, a state-of-the-art real-time detector. We also isolate the reconstruction loss gradient from the detection pathway, ensuring that the backbone and neck are optimized solely by the detection loss. Based on experiments using the VisDrone2019 dataset, we demonstrate that the proposed method achieves respective 𝑚𝐴𝑃50and 𝑚𝐴𝑃50:95 improvements over YOLOv26-N equal to 13.04% and 7.20%, while reducing inference time by 41.3% compared with that of the model integrating the existing self-reconstruction-based approach with comparable detection performance. Our proposed method provides a more efficient and suitable framework for integrating self-reconstruction-based feature enhancement into real-time object detectors, enabling its benefits to be utilized without substantial computational overhead