FCHANet: Feature Cross Hierarchical Fusion Attention Network for Object Detection

Object detection is a fundamental task in computer vision, and deep learning has led to significant advancements in recent years. However, existing methods still struggle with accurate localization and recognition, particularly when detecting objects across multiple scales and varying appearances. These challenges are primarily due to limitations in feature aggregation and fusion techniques. To address these issues, FCHA-Net is proposed, a novel network that utilizes Biformer as the backbone and incorporates two key modules, HAM and FCAF, at the neck. The HAM module enhances multi-scale feature aggregation, enabling better capture of object details across different sizes. The FCAF module applies a cross-attention mechanism to fuse multi-scale features, preserving fine-grained information and improving the accuracy of feature fusion. The combination of these two modules provides an effective solution to the multi-scale detection problem. Training conducted with the MMDetection framework on the PASCAL VOC dataset highlights that FCHA-Net achieves a 93.2% mean average precision (mAP).Ablation studies further validate the effectiveness of the HAM and FCAF modules, demonstrating FCHA-Net’s strength in multi-scale feature fusion and information retention. These results introduce a promising approach to improving object detection and offer new insights for future advancements in the field.