【摘要】 针对声呐图像的目标检测任务，提出一种改进型水下声呐图像目标检测算法。该算法在噪声较大和目标信息缺失的水下声呐图像上表现了出色的检测性能。引入了应用于极端天气下的图像增强模块GDIP，并将其改进为SDIP(Sonar DIP)模块。改进后的SDIP模块应用在声呐图像上效果显著，能够自适应调整增强算法参数，提高了模型的检测精度。还改进了像素级增强阈值，增强图像的同时，抑制了噪声的增强，同样将该阈值设为可优化参数，通过模型训练来自适应调整。检测模型采用CNN与Transformer融合结构，围绕DCNv3算子搭建Transformer网络中更高效的Layer Normalization(LN)、前馈神经网络FFN等结构。优化了模型的内部架构，通过加入EC、Hdc、PC模块来提升模型的性能，其原理是基于轻量级深度卷积进行构建，捕获局部空间信息，对噪声较大和信息缺失严重的水下声呐图像具有显著的效果。实验结果表明，加强空间信息的获取能有效提高模型对该任务的检测性能。更多还原

【Abstract】 An improved target detection algorithm for underwater sonar images is proposed. The proposed algorithm achieves excellent detection performance on underwater sonar images with high noise and incomplete target details. The image enhancement module GDIP, which is applied in extreme weather, is introduced and improved to SDIP(Sonar DIP) module, and the improved SDIP module is applied to sonar images with remarkable effect, which can adaptively adjust the parameters of the enhancement algorithm to improve the detection accuracy of the model. Meanwhile, a pixel-level enhancement threshold is also added to enhance the image details while suppressing noise amplification, which is also set as an optimizable parameter to be self-adaptively adjusted through model training. The detection model adopts the fusion structure of CNN and Transformer, and construts the more efficient and advanced Layer Normalization(LN) and Feed-forward Neural Network(FFN) structures in the Transformer network based on the DCNv3 operator. The internal architecture of the model is optimized, and EC, Hdc, and PC modules are added to enhance the performance of the model, which is based on the principle of lightweight deep convolution construction to capture local spatial information, and it has significant effect on underwater sonar images with large noise and serious information loss. The designed experiments prove that enhancing the acquisition of spatial information can effectively improve the detection performance of the model for this task.更多还原