Small object detection based on hierarchical attention mechanism and multi‐scale separable detection

• Published in: IET image processing Vol. 17; no. 14; pp. 3986 - 3999
• Main Authors: Zhang, Yafeng, Yu, Junyang, Wang, Yuanyuan, Tang, Shuang, Li, Han, Xin, Zhiyi, Wang, Chaoyi, Zhao, Ziming
• Format: Journal Article
• Language: English
• Published: Wiley 01.12.2023
• Subjects: image recognition; object detection
• ISSN: 1751-9659; 1751-9667
• DOI: 10.1049/ipr2.12912

The ability of modern detectors to detect small targets is still an unresolved topic compared to their capability of detecting medium and large targets in the field of object detection. Accurately detecting and identifying small objects in the real‐world scenario suffer from sub‐optimal performance due to various factors such as small target size, complex background, variability in illumination, occlusions, and target distortion. Here, a small object detection method for complex traffic scenarios named deformable local and global attention (DLGADet) is proposed, which seamlessly merges the ability of hierarchical attention mechanisms (HAMs) with the versatility of deformable multi‐scale feature fusion, effectively improving recognition and detection performance. First, DLGADet introduces the combination of multi‐scale separable detection and multi‐scale feature fusion mechanism to obtain richer contextual information for feature fusion while solving the misalignment problem of classification and localisation tasks. Second, a deformation feature extraction module (DFEM) is designed to address the deformation of objects. Finally, a HAM combining global and local attention mechanisms is designed to obtain discriminative features from complex backgrounds. Extensive experiments on three datasets demonstrate the effectiveness of the proposed methods. Code is available at https://github.com/ACAMPUS/DLGADet Here, we present a novel small object detection method, called deformable local and global attention, which utilizes the hierarchical attention mechanisms and deformable multi‐scale feature fusion to enhance the recognition and detection performance of small object. Our research addresses the key challenges of small target size, complex background, variability in illumination, occlusions, and target distortion that often degrade the performance of small traffic sign detection algorithms.