改进YOLOv5框架在细菌计数方向的研究

Improved YOLOv5 Framework in the Direction of Bacterial Counting

摘要: 水中细菌数量的多少作为衡量水质的重要指标之一，其变化可以间接地反映水体污染程度。水中细菌总数反映了水体被有机物污染的程度。为了能快速、高效、准确地统计出细菌总量，将深度学习引入环境工程中，提出一种基于YOLOv5的细菌计数改进算法。首先对自建细菌数据集使用K-means++聚类算法获得和特征图更加匹配的先验框，之后，在网络中增加一层小目标检测层，提高模型对图像中小目标的敏感度，最后，在骨干网络中C3层后引入一种协调注意力(CA)，其不仅能捕获跨通道信息，还能捕获方向感知和位置敏感信息，提高对小目标的识别度，这有助于模型提高对密集预测任务的性能。实验表明，相比于传统的YOLOv5框架算法，改进后的算法在测试集上的平均检测率到达93.04%，提高了7.68%，同时训练损失也更低，验证了增加小目标检测层和注意力机制对细菌图像这种小目标密集检测有较好效果。该算法的引入可以提高细菌计数效率和计数精准度，同时实现对细菌数量的高精度分析，从而进一步深入研究微生物群落的结构、环境污染的程度以及疾病的诊断与治疗等方面，为环境监测提供了有力支持。

Abstract: The number of bacteria in water is one of the important indicators to measure water quality, and the change of bacteria number can indirectly reflect the degree of water pollution. At the same time, the total number of bacteria in water reflects the degree of pollution by organic matter. In order to count the total amount of bacteria quickly, efficiently and accurately, an improved algorithm of bacteria counting based on YOLOv5 is proposed by introducing deep learning into environmental engineering. Firstly, a K-means++ clustering algorithm is used for the self-built bacteria dataset to obtain priori frames that match more closely with the feature map. Secondly, a small target detection layer is added to the network to improve the sensitivity of the model to small targets in images, finally, a coordinated attention (CA) is introduced after the C3 layer in the backbone network, which can capture not only cross-channel information but also orientation-aware and position-sensitive information to improve the recognition of small targets, which helps the model to improve its performance for dense prediction tasks. Experiments show that the improved algorithm achieves an average detection rate of 93.04% on the test set compared to the traditional YOLOv5 framework algorithm, an improvement of 7.68%, as well as a lower training loss, verifying that the addition of the small target detection layer and the attention mechanism is more effective for dense detection of small targets like bacterial images. The introduction of this algorithm can improve the efficiency and accuracy of bacterial counting, and can achieve high precision analysis of bacterial counts, further deepening the study of the structure of microbial communities, the degree of environmental pollution, and the diagnosis and treatment of diseases, providing strong support for environmental monitoring.