Small Object Detection with Super-Resolution Cascaded YOLOv4: An Average Precision Analysis

Object detection is a fundamental aspect of computer vision, especially for satellite image analysis. Detecting small objects remains challenging due to limited resolution and features distortion, impede deep learning models’ learning ultimately impacting detection performance. To address this challenge, this paper investigates regression-based super-resolution (SR) models to generate high-resolution satellite images, aiming to enhance small object clarity and detail, and focus on quantitatively evaluate the influence of SR on average precision (AP) of small object detection. For this, we propose a deep learning pipeline, cascades regression-based SR architectures, including Super-Resolution Convolutional Neural Network (SRCNN), Very Deep Super-Resolution (VDSR) network, and Residual Channel Attention Network (RCAN), individually with the YOLOv4 to detect the small objects, particularly the "Planes" class in the DOTA v1.5 dataset. Experimental results indicate that SRCNN3(3-layer) outperforms VDSR20 and RCAN variants (RCAN20 & RCAN100) in mean BRISQUE scores but does not improve YOLOv4 detection when cascaded, likely due to alters textures and edges in manner that is suboptimal for YOLOv4-based detection. Conversely, RCAN100 demonstrate encouraging results, potentially improving detection performance, despite lower BRISQUE scores than SRCNN3. This indicates that cascading regression-based SR does not necessarily guarantee to enhance YOLOv4-based object detection performance. The source code is available at https://github.com/SharazJaved/SR_YOLO_DOTA.