基于多级特征混叠融合的水下小目标检测算法

An Multilevel Feature Hybrid Fusion Algorithm of Underwater Small Object Detection

针对水下探测机器人在动态巡检过程中其检测的目标尺度存在多样化变化,而传统的SSD算法在尺度特征提取能力方面有限,对远目标、小目标存在信息感知偏弱问题,提出一种基于多级特征混合融合的深度学习目标检测算法。该算法以SSD算法的骨干网络为基础,引入卷积注意力机制增强浅层网络的特征提取能力,同时,提出融合多层高语义特征信息与低语义特征信息的级联式混合融合结构,通过混合训练解决低语义层级对小目标信息特征提取困难的问题。利用水下机器人目标抓取大赛的数据集对算法的有效性进行验证,实验结果表明该算法较传统SSD算法检测精度提升了5.86%。

In the process of dynamic inspection,the traditional SSD algorithm is limited in the scale feature extraction ability and has weak information perception for remote and small targets.For the above situation,a deep learning object detection algorithm structure based on multi-level feature hybrid fusion was proposed in this paper.Based on the backbone network of SSD algorithm,the convolutional attention mechanism was introduced to enhance the feature extraction ability of shallow network,at the same time,the different levels of high semantic feature information and low semantic feature information used cascade structure for multi-level fusion,through mixed training low semantic level of small target information feature extraction problem was solved.This paper verified the effectiveness of the algorithm,and the experimental results showed that the detection accuracy was 5.86%compared to the traditional SSD algorithm.