A lightweight object detection method based on fine-grained information extraction and exchange in UAV aerial images

Objects in unmanned aerial vehicle (UAV) images are easily disturbed by complex backgrounds, and objects different positions in these images often have notable differences in size because of different shooting angles. To effectively detect objects, current mainstream methods improve the detection accuracy through complex iterative convolution operations and attention mechanisms. However, these methods not only improve the accuracy but also result in a high memory overhead and feature redundancy, which brings unbearable load pressure to the UAV platform. Therefore, to refine the multi-granularity object information of UAV aerial images via a lightweight manner and improve the precision of multi-scale object detection, we design an portable lightweight multi-scale UAV image object detection network (UAVDNet) based on MConvBottleNet. First, to overcome the challenge that the conventional convolution receptive field is unitary and easily loses the fine-grained information described above, we design a multifunctional convolution (MConv) module achieve multi-receptive field information aggregation and feature weighting through hierarchical mechanism. Second, we propose MConvBottleNet to simultaneously aggregate local and global information using residual connections and channel shuffling operations on the basis of the diverse information provided by MConv. Third, to effectively exploit the context information in high-level semantic feature maps and preserve the original fine-grained details to the maximum possible extent, we design an inter-layer cascaded information aggregation pooling (ICIAP) module, which, together with MConvBottleNet, constitutes the feature extraction network. Finally, we propose a fusion network based on the feature recombination and enhancement (FRE) module, denoted as FRENet, which can take advantage of the information-complementary characteristic different channel layers to obtain overall channel enhancement results and effectively improve the ability to detect multi-scale objects. Experiments on the VisDrone dataset show that UAVDNet achieves an average detection accuracy of 48.1% with only 4.4M parameters.