通道空间深度感知的轻量化水下目标检测

Lightweight underwater target detection for channel spatial depth perception

摘要:

提出了一种通道空间深度感知的轻量化水下目标检测网络CSDP-L-YOLO。该网络基于YOLOv5网络进行改进，由特征感知模块和双注意门控策略组成。特征感知模块旨在将解码器中的多级特征自适应抑制或增强，优化类内学习的一致性，解决水下场景复杂导致的误检和漏检问题；通过线性操作和混洗结构生成特征映射，减少冗余特征的融合和计算，以减少模型的参数量和计算量。双注意门控策略是在编码器中同时引入并发通道空间挤压-激励机制模块和卷积注意力模块，进一步关注强相关性特征，增强模型对特征的敏感度。实验结果表明，与基线模型YOLOv5-s相比，mAP提高了2.4%，节省了20%参数量和15.8%计算量，检测速度提升了8.2 ms。此外，与目前较为先进的YOLOv8模型相比，mAP提高了1.9%。

Abstract:

A lightweight underwater target detection network CSDP-L-YOLO for channel spatial depth perception is proposed. The network is improved based on the YOLOv5 network and consists of a feature awareness module and a two-attention gating strategy. The feature sensing module aims at adaptive suppression or enhancement of multi-level features in the decoder, optimizing the consistency of in-class learning, and solving the problem of false detection and missing detection caused by the complexity of underwater scenes. The feature mapping is generated by linear operation and mixing structure to reduce the fusion and calculation of redundant features, so as to reduce the number of parameters and calculation amount of the model. The dual attention gating strategy is to introduce concurrent channel space squeezing and stimulation module and convolutional attention module into the encoder at the same time to further focus on the strong correlation features and enhance the sensitivity of the model to the features. The experimental results show that compared with the baseline model, mAP improves by 2.4%, saves 20% parameters and 15.8% computation, and improves the detection speed by 8.2 ms. In addition, mAP improves by 1.9% compared to the current more advanced YOLOv8 model.