Note: Some of the code comes from https://github.com/bubbliiiing/yolox-pytorch.
#The innovation part has shown effective improvements on the DIOR dataset, and a paper is currently under review. #If it is accepted, we will share the link to the paper in the future. #Description of Innovations:
The work of this paper mainly consists of four parts, namely: Additional Predictive Feature Layer module (APFL), Fusion Standard Deviation Feature Enhancement Module (FSEM), Adaptive Feature Fusion Module (AFM), and an improved loss function.
APFL: The proposed idea of APFL is as follows: YOLOX-Tiny cannot effectively cover all objects, especially dense and proportionally different remote sensing objects. For a remote sensing image with an input resolution of 416×416, the model uses only three scales of feature maps for prediction, which are 13×13, 26×26, and 52×52, and can effectively recognize objects larger than 32 pixels × 32 pixels, 16 pixels × 16 pixels, and 8×8 pixels. However, for objects smaller than 8 pixels × 8 pixels, the detection loss probability is relatively high. Therefore, we added the APFL module to improve the detection accuracy of small targets. The Improved algorithm ablation experiments in section 18 of the paper show that APFL improved the detection accuracy of targets by 0.52%.
FSEM: Attention mechanisms can make it easier for networks to focus on targets in the background. However, most attention mechanisms currently used, such as SENet, CBAM, and ECANet, use global average pooling or global maximum pooling to replace image features as the learning prerequisite for attention coefficients, ignoring the prerequisite of global average pooling or global maximum pooling as the learning prerequisite for attention coefficients, resulting in poor performance of attention mechanisms in complex background remote sensing images. Based on this, we designed the FSEM module, which integrates the standard deviation values and the attention coefficients of the entire dataset to effectively focus on the targets in remote sensing images containing noise information. The effectiveness of the FSEM module was verified in the Improved algorithm ablation experiments section on page 18 of the paper.
AFM:The well-known feature pyramid network FPN and its variants fuse shallow and deep features. However, this fusion method is the same and static for the entire dataset, ignoring the feature fusion considering the specific input image, which undoubtedly limits the adaptiveness of feature fusion and affects the fusion effect. To address this issue, we propose the AFM module, combining the idea of residual network, which can perform multi-scale feature adaptive fusion according to the different features of specific input images. The effectiveness of the AFM module is validated in the Improved algorithm ablation experiments section on page 18 of the paper.
Loss function: To improve the misalignment problem between the decoupled regression and classification tasks in the YOLOX-Tiny algorithm, this paper modified the original algorithm's loss function. The modified loss function fuses the predicted confidence values and the predicted classification results, enabling the classification and regression tasks to interact more effectively during network training. The effectiveness of the modified loss function was verified in the Improved algorithm ablation experiments section on page 18 of the paper