Confidence-Based Weakly Supervised for Aircraft Detection From Remote Sensing Image

Weakly supervised aircraft detection is a crucial topic in object detection. The existing approach uses a class activation map (CAM) to generate a pseudo-label. However, the reliability of the pseudo-label is often overlooked as current methods mainly focus on expanding the activating regions. Weak supervision relies on confidence to determine pseudo-labeling reliability. Pseudo-label confidence has often been measured by the peak values of the predicted probability distribution, which is essentially a bootstrap method that is not objective enough to provide precise predictions. In this letter, we propose a confidence-based weakly supervised framework (CWSF) to improve the reliability of pseudo-labels. A multiscale feature fusion and refinement module (MFFR) is designed to optimize and extract complex features in remote sensing images. We also introduce a confidence predictor (CP) to determine pseudo-label confidence. Moreover, the loss function (CWSLoss) is proposed to train the network framework. The proposed CWSF improves the quality of pseudo-labeling and object detection results. Our experiments on selected datasets NPWU VHR-10 (67.01%) and PatternNet (65.42%) show that CWSF significantly outperforms previous state-of-the-art methods.