MLA-Net: Feature Pyramid Network with Multi-Level Local Attention for Object Detection

Feature pyramid networks and attention mechanisms are the mainstream methods to improve detection performance of many current models. However, when they learned jointly, there is a lack information association between multi-level features. Therefore, this paper proposes feature local method, dubbed as MLA-Net (Feature Pyramid Network with Multi-Level Local Attention for Object Detection), which aims establish correlation mechanism information. First, original features deformed rectified using pixel-rectification module, global semantic enhancement achieved through spatial-attention module. After that, further fused residual connection achieve fusion contextual enhance representation. Extensive ablation experiments were conducted on MS COCO (Microsoft Common Objects in Context) dataset, results demonstrate effectiveness proposed method 0.5% enhancement. An improvement 1.2% was obtained PASCAL VOC (Pattern Analysis Statistical Modelling Computational Learning, Visual Classes) reaching 81.8%, thereby, indicating that robust can compete other advanced