基于残差连接的水下小目标检测结构模型

Underwater Small Object Detection Model Based on Residual Connections

由于水下成像距离较远,现有的水下图像目标检测数据集中存在大量小目标,水下小目标特征信息和语义信息少,目前常用的目标检测算法直接应用于水下检测时小目标漏检错检率较高。水下目标检测预处理阶段通常采用图像增强算法来提升图像观感质量,并以此提升小目标检测精度,但图像增强的数据预处理操作容易导致小目标特征丢失,小目标检测性能下降。提出了一种基于残差连接的水下小目标检测结构模型,讨论了图像增强提升目标检测性能的应用方式,通过残差连接将增强算法与目标检测算法联合优化,避免了过度增强导致的特征丢失问题,提升了水下小目标检测精度。所提出的算法在DUO数据集上进行了实验,实验结果表明,相较于YOLOv7,该算法对小目标检测精度提升了7.2%。在两个水下数据集上进行消融实验,验证了所提出的残差连接的方式对于提升小目标检测性能具有促进作用。

Due to the long distance of underwater imaging,there are a large number of small objects in the existing underwater image object detection datasets,and there is little feature and semantic information of underwater small objects.When the existing object detection algorithms are directly applied to underwater detection,the missed detection and false detection rate of small objects is relatively high.In the preprocessing stage of underwater object detection,image enhancement algorithms are usually used to improve the visual quality of images and improve the accuracy of small object detection.However,the data preprocessing operation of image enhancement can easily lead to the loss of small object features and a decrease in the performance of small object detection.Therefore,a structural model for underwater small object detection based on residual connections is proposed,and the application methods of image enhancement to improve object detection performance are discussed.The enhancement algorithm and object detection algorithm are jointly optimized through residual connections,avoiding the problem of feature loss caused by excessive enhancement and improving the accuracy of underwater small object detection.The proposed algorithm was tested on the DUO dataset,and the experimental results showed that compared to YOLOv7,the algorithm improved the accuracy of small object detection by 7.2%.Ablation experiments were conducted on two underwater datasets,verifying that the proposed residual connection method has a promoting effect on improving the performance of small object detection.