Noise-Resilient Small Object Detection with Octave and Cross-Frequency Convolution

Detecting small objects in noisy environments is difficult due to their limited size, complex backgrounds, and the degradation caused by various noise sources. To address these challenges, this work presents a noise-resilient object detection framework based on a latest You Only Look Once (YOLO) architecture. The framework includes an auxiliary neck using Octave Convolution to separate high-frequency components, which are essential for small object detection, from low-frequency redundant information. A Noise-Resilient Detection Block further improves performance by enhancing fine details while reducing the impact of noise. In addition, the integration of Space-to-Depth Convolutions and Convolutional Block Attention Modules preserves fine-grained spatial details and strengthens the focus on important regions. Evaluations on the real-world datasets demonstrate that the proposed framework outperforms leading existing methods, while its lightweight and efficient design makes it suitable for real-time and embedded applications.