基于残差注意力网络模型的浮游植物识别

浮游植物作为水生态系统中最重要的生物组成部分之一，对水环境敏感，在水环境监测中得到了广泛的关注。然而水生环境复杂多样，准确高效地识别浮游植物是监测工作中的一大挑战。当前浮游植物识别方法可分为经典形态学分类、分子标记和人工智能图像识别三类。前两种方法已被广泛采用，但费时费力，不利于监测机构的大规模应用和推广。同样，利用图像进行自动化分类难以在高准确率与高效率上达到平衡。深度学习技术的发展为此提供了新思路。本文提出一种新的深度卷积神经网络RAN-11。该网络以残差注意力网络Attention-56和Attention-92为基础，凭借通道对齐融合主干上的底层特征与顶层特征，通过调整注意力模块和残差快个数以精简结构，并引入了Leaky ReLU激活函数代替ReLU。以太湖11个优势属共计1036张图像为数据来源进行对比验证。除星杆藻外，RAN-11对单一优势属的的查准率都在90%以上，并且有5个优势属达到100%的查准率。RAN-11的识别准确率为95.67%，推理速率为41.5帧/s，不仅比Attention-92（95.19%的准确率，23.6帧/s）更准确，而且比Attention-56（94.71%的准确率，41.2帧/s）更快，真正兼顾了准确率与效率。研究结果表明：（1）RAN-11在查准率、准确率和推理速率上优于原始残差注意力网络，更优于以词包模型为代表的传统图像识别方法；（2）融合多尺度特征、精简网络结构和优化激活函数是提高卷积神经网络性能的有力手段。建立在经典分类基础之上，本文提出新的残差注意力网络来提升浮游植物鉴定技术，并构建出浮游植物自动化识别系统，识别准确率高、易于推广，对于实现水体中浮游植物的自动化监测具有重要意义。

Phytoplankton recognition based on residual attention network

Phytoplankton, one of the most important biological components in water ecosystem, is sensitive to water environment and has been widely concerned in water environment monitoring. However, the aquatic environment is complex and diverse. Accurate and efficient identification of phytoplankton becomes a challenge in monitoring work. Current phytoplankton recognition methods can be divided into three categories:classical morphological classification, molecular markers, and artificial intelligence image recognition. Although the first two methods have been widely used, they are time-consuming and laborious, which is not conducive to the large-scale application and promotion of monitoring agencies. Similarly, it is difficult to strike a balance between high accuracy and high efficiency for automatic classification based on images. Deep learning provides new insights for phytoplankton identification. We propose a new deep convolutional neural network RAN-11. The network is based on the residual attention network Attention-56 and Attention-92, fusing the bottom and top features of the backbone through channel concatenation, simplifying the structure by adjusting the number of attention and residual module, and adopting the Leaky ReLU activation function rather than ReLU. We used 1036 images of 11 dominant genera of Taihu Lake as data sources to compare and verify our algorithm. The precision of RAN-11 for a single prevailing genus was above 90%, with 5 species achieving 100% precision, except for Asterionella sp. The accuracy of RAN-11 was 95.67%, and the inference speed was 41.5 fps (frames per second), which is not only more accurate than Attention-92 (95.19% accuracy, 23.6 fps), but also faster than Attention-56 (94.71% accuracy, 41.2 fps), truly balancing accuracy and efficiency. Results indicate that:(1) RAN-11 is superior to the original residual attention network in terms of precision, accuracy and inference speed, as well as the traditional image recognition method represented by the Bags of Words model. (2) Fusion of multi-scale features, simplification of network structure and optimization of activation function are powerful means to improve the performance of convolutional neural network. Based on the classical classification, this paper proposes a new residual attention network to improve phytoplankton identification technology, and constructs an automatic phytoplankton recognition system, which has high recognition accuracy and easy promotion, and is of great significance to realize the automatic monitoring of phytoplankton in water.