结合双线性网络和注意力的蓝藻图像识别方法

Image recognition method for cyanobacteria utilizing bilinear networks and attention mechanisms

摘要:

针对实际工程环境中采集的蓝藻图像数量不均衡、复杂光照条件影响、蓝藻图像局部区域特征捕捉不全面等问题，提出一种结合双线性网络和注意力的蓝藻图像识别方法。首先构建蓝藻图像数据集，并使用图像增强算法对数据集进行优化，再使用双线性网络充分提取蓝藻图像特征信息，同时结合卷积块注意力机制，聚焦重要的局部特征，忽视无用信息，进一步提升对无蓝藻、颗粒状蓝藻、带状蓝藻和片状蓝藻等 4 类不同形态蓝藻图像的分类效果。在构建的 Algea-ultimate 蓝藻数据集上进行实验，结果表明：相比识别效果最好的经典网络模型 ResNet18，所提方法在识别准确率上提升 7.29%，识别精度有明显提升。识别方法可用于太湖流域水质监测和预警平台中，提供蓝藻图像自动识别功能，为实时监测水体蓝藻形态提供智能化解决方案。

Abstract:

Addressing issues such as the uneven quantity of cyanobacteria images collected in practical

engineering environments，the influence of complex lighting conditions and the incomplete capture of local

features in cyanobacteria images，a cyanobacteria image recognition method combining bilinear networks and

attention mechanisms is proposed. Firstly，a cyanobacteria image dataset is developed and optimized using an

image enhancement algorithm. Subsequently，bilinear networks are employed to comprehensively extract feature

information from cyanobacteria images. Simultaneously，the convolutional block attention mechanism is integrated

to emphasize important local features while disregarding irrelevant information. This approach aims to further

improve the classification performance across four distinct types of cyanobacteria image including cyanobacteriafree，

granular cyanobacteria，banded cyanobacteria and flake cyanobacteria. Experimental results conducted on

the constructed Algea-ultimate cyanobacteria dataset demonstrate a notable enhancement in recognition accuracy

compared to the classical ResNet18 model，with a 7.29% improvement. Furthermore，the proposed method has been

implemented in the Taihu Lake Basin water quality monitoring and early warning platform to facilitate automated

cyanobacteria image recognition，off ering an intelligent solution for real-time monitoring of water body cyanobacteria

morphology.