DECOMPOSABLE MULTI-KERNEL CONVOLUTIONAL NETWORKS FOR OBJECT DETECTION IN REMOTE SENSING IMAGES

Remote sensing object detection (RSOD) algorithms can accurately capture the distribution of objects on the Earth’s surface with powerful spatial resolution. Deep learning-based RSOD algorithms have achieved excellent performance, which is mainly attributed to strong backbone networks with powerful feature extraction abilities. However, existing backbone networks typically use fixed receptive fields to extract image features, neglecting the requirement for varying receptive fields in RSOD. Additionally, the large number of parameters and high computational costs of these models limit their real-time performance. In this paper, a Decomposable Multi-kernel Convolutional Network containing a Multi-kernel Receptive Field Module (MRF) and a Feature Selection Module is proposed for RSOD. Specifically, the MRF module employs sparse convolutional computations to achieve feature extraction across multiple receptive fields, exploring the local spatial features of targets at multiple scales while reducing model complexity. Meanwhile, the FSM module leverages attention modeling in both spatial and channel dimensions for feature selection, enhancing effective object features across different receptive fields. The proposed algorithm is evaluated on the publicly available DIOR and DOTA datasets, and the experimental results show that our method significantly outperforms the other state-of-the-art methods.