注意力机制优化的全尺寸目标检测方法

针对现有目标检测算法全尺寸目标检测精度低的问题，提出了一种改进的基于YOLOv3模型的全尺寸目标检测算法。该方法设计了一种全新的通道自适应递归FPN网络架构，提出了一种基于通道注意力的递归金字塔模型，提高了YOLOv3的特征提取能力和不同尺度目标的检测能力。同时在训练过程中引入损失函数转换，解决了训练过程中动态参数不优化的问题。与其他主流目标检测算法相比，本文提出的改进模型在小尺寸目标、大尺寸目标与复杂背景多尺寸目标的检测精度分别提高了5.6%、2.6%、1.6%。实验结果表明，本文提出的方法检测精度显著提升。

Full-size object detection method optimized by attention mechanism

Aiming at the problem that existing object detection algorithms have low accuracy in full-size object detection, this paper proposes an improved full-size object detection algorithm based on the YOLOv3 model. In the method, a new adaptive recursive FPN network architecture is designed, and a recursive pyramid model based on channel attention is proposed to improve the feature extraction ability of YOLOv3 and the detection ability of objects at different scales. At the same time, loss function transformation is introduced in the training process to solve the problem of dynamic parameters that is not being optimized in the training process. Compared with other mainstream object detection algorithms, the accuracy of small-size objects, large-size objects and multi-size objects with complex backgrounds respectively improved by 5. 6%, 2. 6%, and 1. 6%. Experimental results show that the detection accuracy of the proposed method is significantly improved.