Efficient Enhanced Feature Learning for Remote Sensor Image Object Detection

The realm of computer vision has witnessed a profound interest from the scholarly community in the task of remote sensing object detection. Within this article, we present a pioneering and refined approach to feature learning, wherein we harness the potential of the backbone and neck network. We have devised a sophisticated threshold network for the backbone network, facilitating the extraction of profound insights and bolstering our capacity to acquire features by effectively addressing the challenge of high-noise perturbations present in remote sensing imagery. We propose an upgraded feature network paradigm for the neck network, primarily consisting of the Adaptive Feature Enhanced Module (AFEM) and the Gated Convolution Module (GCM). The AFEM is designed to capture contextual information at larger scales, extending the receptive field and preventing feature resolution degradation. On the other hand, the GCM utilizes gated convolution with high-order spatial interactions to adaptively reduce superfluous semantic information. We have undertaken comprehensive experiments on remote sensing image datasets, including DOTA and HRSC, showcasing the exceptional performance of our approach.