AB-YOLO: A Novel Deep Learning Approach for Accurate and Efficient SAR Target Detection

In this paper, we introduce an innovative Synthetic Aperture Radar target detection approach that mitigates the impact of noise on radar detection by leveraging an AB-YOLO algorithm. The proposed method integrates an adaptive kernel convolution module, AKconv, which dynamically adjusts sampling shapes to effectively capture features of diverse forms and locations. This is followed by a dual-layer routing attention mechanism that enhances dynamic sparse attention, filtering irrelevant key-value pairs at a coarse level and preserving only the most pertinent routing areas to reduce redundancy. We further apply a Token-to-Token attention mechanism within these joint routing areas, linearly mapping feature maps and utilizing deep convolutional computation with the most relevant token regions as keys to bolster local feature extraction. The resultant local features of varying scales are sampled, fused, and mapped to generate detection outcomes. Extensive experimental validation on the SAR-AIRcraft-1.0 dataset confirms the superior detection capabilities of our proposed method, showcasing its robustness and precision in noisy conditions.