基于联合深度统计特征对齐的鱼类目标识别方法

Fish target recognition method based on joint deep statistical feature alignment

水下鱼类目标识别技术是认识海洋、经略海洋、向海图强的重要技术之一.基于深度学习的水下目标识别技术已成为研究热点,但是针对水下鱼类数据小样本甚至零样本识别性能亟待提高.本文基于迁移学习,提出了联合深度统计特征对齐(Joint Deep Statistical Feature Alignment, JDSFA)方法,解决小样本下的鱼类目标识别问题.以ResNet-50作为骨干网络,将均方和协方差纳入权重选择算法用来构建自适应损失函数,对齐源域和目标域之间的特征分布,联合源域损失与领域间的自适应损失,设计全局损失函数,建立深度学习识别模型,实现鱼类目标识别任务.利用公开的水下鱼类数据集QUT进行实验验证,相比目前代表性的DADAN、PMTrans、DSAN方法,JDSFA方法的鱼类识别性能分别提升了3.59%、4.96%、5.91%,结果表明了本文JDSFA方法的有效性,并对鱼类目标识别具有良好的应用价值.

Underwater fish target recognition technology is an important technology in ocean exploration.However,the acquisition of underwater resources is very expensive and cannot meet the requirements of training large-scale deep neural networks.Aiming at the problem of small or even zero samples of underwater fish data,this paper proposed a Joint Deep Statistical Feature Alignment(JDSFA)method based on transfer learning to solve the problem of fish target identification in small samples.Using ResNet-50 as the backbone network,the mean square and covariance were incorporated into the weight selection algorithm to construct a joint domain adaptive loss function to align the feature distribution between the source domain and the target domain.In addition,we also designed a global loss function to combine the source domain loss with the inter-domain adaptive loss by using adaptive weights to realize the fish target recognition task.The published underwater fish data set QUT was used for experimental verification,and the results all showed that the effectiveness of JDSFA method in fish identification was 3.59%,4.96%and 5.91%higher than that of the current representative DADAN,PMTrans and DSAN methods,respectively.It is further demonstrated that the JDSFA method proposed in this paper has good application value in fish identification task.