A Small Target Detection Method based on Multibranch Convolution and Adaptive Dual Attention Module

Despite significant advances in object detection, existing convolutional and attention modules still struggle to focus effectively on small object regions and extract their fine-grained features, leading to detection performance that falls short of practical requirements. To address this issue, this paper proposes two modules: the Multibranch Convolution Module (MConv) and the Adaptive Dual Attention (ADA) Module. Specifically, MConv decomposes the feature map into four sub-maps via pixel-level slicing to preserve spatial information without down-sampling. It then applies 3 × 3 convolutions independently to each sub-map to capture local context and inter-pixel relationships, which are critical for small objects. Additionally, residual connections are introduced to ensure gradient flow and retain fine-grained details, further enhancing feature extraction capabilities. ADA is lightweight and plug-and-play. It generates attention maps in parallel across both channel and spatial dimensions through input-dependent weight learning, where attention weights are dynamically derived from the local and global statistics of the input feature map, thereby enhancing the capture of small-object features. We validated our approach on five public datasets across four different domains. Experimental results demonstrate that both proposed modules effectively improve small-object detection performance and offer significant practical value. The code is available at: https://github.com/Allisway/ADA-MC.