基于特征融合与通道注意力的目标检测研究

目标检测作为计算机视觉领域的重要研究分支,受到了广泛关注。目前,特征融合已成为提高目标检测准确率的重要方法,基于特征金字塔网络（FPN）的特征融合方法结合了多维度感受野来弥补信息丢失的思想,改进了FPN,并取得了良好的效果。在众多以FPN为基础结构的特征金字塔模型中,BiFPN既包含了级联信息传递,也包含了跨层特征融合,DyFPN包含了多感受野Inception模块,也包含了动态门机制。受BiFPN和DyFPN的启发,文中提出了一种新的基于FPN且包含注意力机制的特征金字塔网络结构——CAI-BiFPN。CAI-BiFPN沿用了Inception-FPN的思想,在BiFPN的基础上加入了Inception模块,并引入了通道注意力和SE模块。该结构应用了BiFPN的分支注意力与SE模块的通道注意力,将Inception模块放置于BiFPN第4层和第6层,SE模块则放置在BiFPN的第5层。通过一系列简单的改进,相较于BiFPN,CAI-BiFPN的AP_s和AP_l提升了0.7个百分点,AP从31.0%提升到31.3%,提升了0.3个百分点。

Research on Object Detection Based on Feature Fusion and Channel Attention

As an important research branch in the field of computer vision, object detection has received extensive attention. At present, feature fusion has become an important method to improve the accuracy of object detection. Feature fusion methods based on Feature Pyramid Network (FPN) combine the idea of multi-dimensional receptive fields to make up for information loss, improve FPN, and achieve good results. Among many feature pyramid models based on FPN, BiFPN includes both cascaded information transfer and cross-layer feature fusion, DyFPN includes multi-receptive field Inception module, and also includes dynamic gate mechanism. Inspired by BiFPN and DyFPN, this paper proposes a new feature pyramid network structure based on FPN and including an attention mechanism-CAI-BiFPN. CAI-BiFPN follows the idea of Inception-FPN, adds the Inception module on the basis of BiFPN, and introduces the channel attention and SE module. This structure applies the branch attention of BiFPN and the channel attention of the SE module, placing the Inception module in the 4th and 6th layers of BiFPN, and the SE module in the 5th layer of BiFPN. Through a series of simple improvements, the CAI-BiFPN has improved by 0.7% compared to BiFPN. Increased 0.7%, AP increased from 31.0% to 31.3%, an increase of 0.3%.