Abstract:

Many effective techniques for object identification have been proposed as a result of the development of convolutional neural networks (CNNs). Remote sensing object detection (RSOD) is a difficult problem due to the following factors: 1) a complex backdrop of remote sensing images (RSIs) and 2) an excessively uneven size and sparsity distribution of remote sensing objects. Existing approaches are incapable of solving these challenges with high detection accuracy and speed. In this study, we suggest an adaptive balanced network (ABNet) to overcome these challenges. First, we create an improved effective channel attention (EECA) technique to boost the backbone's feature representation capabilities, which may help overcome the challenges of complicated backdrop on foreground items. Then, to capture additional discriminative information, an adaptive feature pyramid network (AFPN) is constructed to aggregate multiscale data adaptively in multiple channels and geographical locations. Furthermore, since the original FPN overlooks rich deep-level characteristics, a context enhancement module (CEM) is suggested to take use of extensive semantic information for multiscale object recognition. Experiment findings on three public datasets show that our technique outperforms the baseline by adding less than 1.5M new parameters

Description:

Adaptive feature pyramid, context exploitation, local cross-channel attention, multiscale object detection, remote sensing image (RSI).