针对已有的目标检测方法在复杂场景中对鱼类目标检测效果不理想的问题,提出了一种基于膨胀卷积和参数重构的鱼类目标实时检测方法.先设计了一种四分支融合卷积结构,在引入少量参数量的情况下,扩大了目标检测的感受野,提升了目标检测的效果.再引入了RepVGG(重构VGG)并联辅助分支思想,在训练过程中使用复杂模型进行特征学习,而在推理过程中对模型中的BN(Batch Normalization)层以及1×1的辅助分支中的参数进行融合,利用参数重构对训练过程的冗余参量进行合并,保证了模型的低参数量和实时推理.基于YOLOv5s进行实验,相比原始的YOLOv5s获得了更高的检测精度和召回率,平均精度(mean Average Precision,mAP)达到83.1%,超越了目前主流的目标检测算法.提出的算法在检测速度上相比原始模型无明显降低,处理速度上达到100FPS,在实现高精度检测的前提下保证了鱼类目标的实时检测,为基于视觉的鱼类检测方案提供了有效的技术支持.
Real-time fish object detection based on dilated convolution and parameter reconstruction
To solve the problem that existing object detection methods are not ideal for fish object detection in complex scenes,a real-time fish object detection method based on expansion convolution and parameter reconstruction is proposed.Firstly,a four-branch fusion convolution structure is designed to expand the perceptual field of object detection and improve the effect of object detection by introducting a small number of parameters.Then a RepVGG(Reconstructed VGG)parallel auxiliary branch idea is introduced to use a complex model for feature learning in the training process,while the parameters in the BN(Batch Normalization)layer of the model and the 1×1 auxiliary branch are fused in the inference process,and the redundant parametric quantities in the training process are merged by using parameter reconstruction to ensure the low parametric number and real-time inference.The experiments are conducted based on YOLOv5s,and higher detection accuracy and recall are obtained compared to the original YOLOv5s,with a mean average precision(mAP)of 83.1%,surpassing the current mainstream object detection algorithms.The proposed algorithm has no significant reduction in detection speed compared with the original model,and the processing speed reaches 100 FPS,which ensures real-time detection of fish object while achieving high accuracy detection,and provides effective technical support for vision-based fish detection solutions.
fish detectioncomputer visionYOLOv5 networkdilated convolutionparameter reconfigurationRepVGG module
国家科技期刊平台
NSTL
万方数据
