基于改进的Mask R-CNN的鱼类识别算法研究

Research on Fish Recognition Algorithm Based on Improved Mask R-CNN

水下鱼类是重要的地球生物资源。针对现有的深度学习模型在水下鱼类图像识别场景中识别效果不佳的问题,提出了一种基于改进的Mask R-CNN的鱼类识别算法模型。首先,采用深度残差网络和特征金字塔结构对水下鱼类图像进行特征提取;其次,选用区域候选网络针对特征图生成感兴趣区域;然后,通过改进的Soft NMS算法对感兴趣区域进行后处理以减少对鱼类目标候选框的误检率;最后,在头部网络中添加级联结构对特征区域进行微调以提升鱼类识别精度。在Fish4knowledge数据集上的对比实验结果表明,改进的鱼类识别算法的平均精度均值为87.4%,相对于基线算法模型精度提升了3.6%。所提算法能够有效提高水下鱼类识别精度,同时减少误检率,提升泛化性能,对我国水下鱼类资源的开发利用具有重要的学术价值和经济价值。

Underwater fishes are important living resources on the earth.Aiming at the poor detection ability of existing deep learning models in underwater fish images detection environments,a fish detection algorithm model based on improved Mask R-CNN is proposed with the use of the state-of-art in deep learning.Firstly,the deep residual network and feature pyramid are used to extract the features of underwater fish images.Secondly,the region proposal network is selected to generate the region of in-terest for the feature maps.Then,the improved Soft NMS algorithm is used to the region of interest that is post-processed to reduce the false detection rate of the fish object candidate frame.Finally,a cascade structure is added to the head network to fine-tune the feature area to improve the accuracy of fish recognition.The results of comparative experiments on the Fish4knowledge dataset show that,the average mean accuracy of the algorithm is 87.4%,which is an improvement of 3.6%compared to the single algorithm mod-el accuracy.The proposed algorithm can effectively improve the recognition accuracy of underwater fish while reducing the false de-tection rate and improving the generalization performance,which has important academic and economic value for the development and utilization of underwater fish resources in my country.