PSC-FPN: Pixel Shuffling Algorithm for Channel Conversion in Feature Pyramid Network

Object detection in remote sensing image is vital in areas like agriculture, urban development, and disaster evaluation. However, these images present substantial challenges because of their directional arbitrariness, large-scale variations, and dense object arrangements. While traditional Feature Pyramid Networks (FPN) have shown progress in multi-scale feature fusion, they suffer from issues such as information loss and aliasing effects. To address these limitations, this paper proposes an improved PSC-FPN network with enhanced information fusion. First, the Pixel Shuffle algorithm and sub-pixel convolution are introduced to design the PSCM module, effectively reducing information loss during channel conversion of high-level feature maps. Second, sub-pixel convolution replaces traditional linear interpolation in the top-down path, enhancing the quality of upsampled feature fusion. Additionally, a bottom-up path inspired by PAFPN strengthens the interaction between features at different scales. Finally, a $\mathbf{3}\times \mathbf{3}$ convolution replaces max pooling to generate additional feature levels, preserving more detailed information. Experimental results on the DOTA-v1.0 and DOTA-v1.5 datasets demonstrate that the improved PSC-FPN achieves significant performance gains, with mAP improvements of 1.42 % and 0.71%, respectively, compared to the baseline FPN model.