基于SPD-Conv结构和NAM注意力机制的鱼群小目标检测

Small Object Detection for Fish Based on SPD-Conv and NAM Attention Module

为解决因水下成像环境退化导致图像分辨率较低,以及因鱼群目标较小等因素导致的检测精度不高的问题,提出了一种结合SPD-Conv结构和NAM注意力机制的改进YOLOv7检测算法。首先,采用Space-to-Depth(SPD)结构改进头部网络,取代了网络中原有的跨步卷积结构,保留了更多的细粒度信息,提升了特征学习的效率,提高了网络对低分辨率图像的检测效果。然后在网络中引入Normalization-based Attention Module(NAM)注意力机制,采用CBAM的模块集成方式,使用BN缩放因子来计算注意力权重,抑制了不显著的特征,提升了小目标检测的准确率。最后针对水下成像退化,对检测图片做反卷积预处理,减小了水下成像退化因素对检测造成的影响。实验结果显示,在WildFish数据集上模型的整体精度达到97.2%,与YOLOv7算法相比提升了7.6%,准确率提升了8.5%,召回率提升了9.8%,与Efficientdet,SSD,YOLOv5及YOLOv8算法相比,所提模型精度分别提升了12.6%,17.8%,4%及2.9%,在Aquarium数据集上模型的整体精度达到80.5%,相比Efficientdet,SSD,YOLOv5,YOLOv7及YOLOv8分别提升了18.4%,11.6%,6.9%,2.0%及2.7%,可以满足水下鱼群识别的需求。

In order to solve the problem of low image resolution due to the degradation of underwater imaging environment and low detection accuracy caused by small fish targets,an improved YOLOv7 detection algorithm combining SPD-Conv structure and NAM attention mechanism is proposed.Firstly,the space-to-fepth(SPD)structure is used to improve the head network,which replaces the original straddle convolution structure in the network,retains more fine-grained information,improves the efficiency of feature learning,and improves the detection effect of the network on low-resolution images.Then,the normalization-based attention module(NAM)attention mechanism is introduced into the network,and the module integration method of CBAM is adopted,and the BN scaling factor is used to calculate the attention weight,which suppresses the insignificant features and improves the accuracy of small target detection.Finally,for underwater imaging degradation,the detection image is deconvolved and preprocessed,which reduces the impact of underwater imaging degradation factors on detection.Experimental results show that in the WildFish dataset,the overall accuracy of the model reaches 97.2%,which is 7.6%higher than that of the YOLOv7 algorithm,the accuracy rate is increased by 8.5%,and the recall rate is increased by 9.8%,compared with the Efficientdet,SSD,YOLOv5 and YOLOv8 algorithms,the accuracy of the proposed model is improved by 12.6%,17.8%,4%and 2.9%,respectively.The overall accuracy of the model reaches 80.5%,which is 18.4%,11.6%,6.9%,2.0%and 2.7%higher than that of Efficientdet,SSD,YOLOv5,YOLOv7 and YOLOv8,respectively,which can meet the needs of underwater fish identification.