基于改进YOLOv4的虾苗智能识别算法研究

Research on Intelligent Recognition and Statistical Algorithm of Shrimps Based on Improved YOLOv4

近年来,虾类养殖发展迅速,已经成为我国水产养殖中的新型支柱产业,给国家及养殖企业带来了巨大的经济与社会价值。在对虾类进行养殖的过程中,对虾苗进行识别和数量统计为虾苗运送销售、科学喂养虾苗、挑选优质虾苗等环节提供了强力支持。因此,对虾苗进行智能识别与统计对虾类养殖业具有很重要的现实意义。但是,因虾苗图像存在反光、虾苗个体较小等特点,对虾苗进行智能识别与统计特别困难。本文针对虾苗识别存在的难点,提出一种基于改进的YOLOv4的虾苗智能识别算法,在原有的YOLOv4基础上,通过聚类方法对虾苗大小进行统计,获得虾苗大小的分布特征。基于该统计结果,对YOLOv4模型进行优化,将预设的8个锚框缩减为4个。为了让网络更好地关注重要特征并抑制不必要的特征,在YOLOv4网络的输出阶段加入SAM模块。实验表明,本文提出的算法在准确率、召回率、m AP值的评价中均取得了最优或次优的结果。

In recent years,shrimp culture has developed rapidly,and has become a new pillar industry in China’s aquaculture,which has brought great economic and social value to the country and aquaculture enterprises.In the process of shrimp culture,the identification and quantity statistics of shrimps provide strong support for the transportation and sales of shrimps,scientific feeding of shrimps,selection of high-quality shrimps and other links.Therefore,the intelligent identification and statistics of shrimps has a very important practical significance for shrimp breeding industry.However,it is very difficult to recognize and count shrimps intelligently because of the reflection and small size of shrimps.Aiming at the difficulties of shrimp identification,this paper proposed an intelligent shrimp identification algorithm based on improved YOLOv4.Based on the original YOLOv4,the size of shrimp was counted by clustering method,and the distribution characteristics of shrimp size were obtained.Based on the statistical results,the YOLOv4 model was optimized,and the preset 8 anchor frames were reduced to 4.In order to make the network pay more attention to the important features and suppress the unnecessary features,the spatial attention module(SAM)was added to the output stage of the YOLOv4 network.Experimental results show that the proposed algorithm achieves the best or suboptimal results in the evaluation of accuracy,recall and mAP value.