随着深度学习技术的发展,水下图像检测近年来受到广泛的关注,为了克服在复杂水下环境下传统小鱼群的误检、漏检和识别准确率低等问题,提出一种改进 YOLOv5 的目标检测方法(INV-YOLOv5).该方法包括将 YOLOv5m 中的 Focus模块替换为卷积模块,提高网络精度;在主干网络(Backbone)中添加多头自注意力机制,增大网络特征提取视野;最后,在网络中引入了内卷算子和加权的特征融合,降低网络的参数量,提高检测精度.在实验阶段,使用 Labeled Fishes in the Wild数据集和 WildFish数据集验证,该方法的平均精度(mAP)分别为 81.7%和 83.6%,与 YOLOv5m 网络相比分别提升了 6%和 14.5%,不仅拥有较高的识别率并且更加轻量化,而且模型大小与 YOLOv5m网络相比减少了 6 M(Mega)左右,验证了所提出的改进方法具有较好的效果.
With the development of deep learning technology,underwater image detection has received extensive attention in recent years.In order to overcome the problems of false detection,missing detection and low detection accuracy of traditional small fish groups in complex underwater environments,an improved YOLOv5 target detection method(INV-YOLOv5)was proposed.The method included replacing the focus module in YOLOv5m with the convolution module to improve the network accuracy.A multi-head self attention mechanism was added to the backbone to increase the view of network feature extraction.Finally,the feature fusion of involution operator and weighting was introduced into the network to reduce the amount of network parameters and improve the detection accuracy.The experimental verification on Labeled Fishes in the Wild dataset and WildFish dataset shows that the average accuracy(mAP)of this method is 81.7%and 83.6%respectively,which is increased by 6%and 14.5%respectively compared with YOLOv5m network.It not only has a higher recognition rate,but also is more lightweight.Compared with YOLOv5m network,the model size is reduced by about 6 M(Mega),which verifies that the proposed improved method has a better effect.