Exploring Underwater Target Detection Algorithm Based on Improved SSD

As the in-depth exploration of oceans continues, the accurate and rapid detection of fish, bionics and other intelligent bodies in an underwater environment is more and more important for improving an underwater defense system. Because of the low accuracy and poor real-time performance of target detection in the complex underwater environment, we propose a target detection algorithm based on the improved SSD. We use the ResNet convolution neural network instead of the VGG convolution neural network of the SSD as the basic network for target detection. In the basic network, the depthwise-separated deformable convolution module proposed in this paper is used to extract the features of an underwater target so as to improve the target detection accuracy and speed in the complex underwater environment. It mainly fuses the depthwise separable convolution when the deformable convolution acquires the offset of a convolution core, thus reducing the number of parameters and achieving the purposes of increasing the speed of the convolution neural network and enhancing its robustness through sparse representation. The experimental results show that, compared with the SSD detection model that uses the ResNet convolution neural network as the basic network, the improved SSD detection model that uses the depthwise-separated deformable convolution module improves the accuracy of underwater target detection by 11 percentage points and reduces the detection time by 3 ms, thus validating the effectiveness of the algorithm proposed in the paper. 随着人类对海洋的不断深入探索，准确、快速地检测水下环境中的鱼类、仿生体及其他智能体对完善水下防御体系显得越来越重要。针对水下复杂环境下目标检测准确率低、实时性差的问题，提出一种基于改进SSD的目标检测算法。该算法用ResNet卷积神经网络代替SSD的VGG卷积神经网络作为目标检测的基础网络，并在基础网络中利用所提出的深度分离可变形卷积模块进行特征提取，提高对水下复杂环境下目标检测的精度及速度。所提出的深度分离可变形卷积主要是在可变形卷积获取卷积核偏移量的过程中融合深度可分离卷积，以减少参数量来达到提升网络运行速度的目的，同时通过稀疏表示来提升网络的鲁棒性。实验结果显示，相比ResNet作为基础网络的SSD检测模型，利用深度分离可变形卷积改进的SSD检测模型检测水下目标的准确率提升了11个百分点，检测时间减少了3 ms，证明新算法的有效性。