基于YOLO的鱼道过鱼粗粒度目标检测算法

YOLO-based coarse-grained object detection algorithm for fish passage in fishways

针对鱼道过鱼目标检测问题中光照多变、水质较差及数据集先验信息不足而导致的漏检,误检问题,文章提出一种基于改进YOLOv5的目标检测算法。该方法首先在预处理部分使用灰度算法将输入的彩色图片转换为灰度图,克服了水下光线变化对识别结果产生的影响;同时在YOLOv5特征提取部分嵌入坐标注意力模块以突出目标特征,在特征融合部分嵌入自适应特征融合模块以提高网络在目标特征模糊条件下的特征融合能力;最后将多类别目标检测简化为单类别目标检测,即粗粒度的鱼类目标检测,以解决数据集先验信息不足问题。该方法应用于真实环境下的鱼道过鱼数据集进行测试,与原始YOLOv5方法相比,提出的方法检测精度达到95%,相比原方法提升了12%。研究成果对鱼道过鱼的目标检测有良好的借鉴意义。

In order to solve the problem of detection of fish crossing targets in fish passages with variable lighting,poor water quality and limitations of the data set,this paper proposes an improved YOLOv5-based object detection algorithm.Firstly,the method uses grayscale algorithm to convert the input color image into grayscale image in the preprocessing part to overcome the effect of underwater light variation on the recognition result;meanwhile,the method embeds coordinate attention module in the feature extraction part of YOLOv5 to highlight the target features,and adaptively spatial feature fusion module in the feature fusion part to improve the feature fusion ability of the network under the fuzzy target features;finally,the multi-category object detection is simplified to single-category object detection,which means coarse-grained fish object detection,in order to solve the problem of data set limitation.The method in this paper is applied to the fish passage over fish dataset in a real environment for testing,and the detection accuracy of the proposed method reaches 95%compared with the original YOLOv5 method,which is 12%better than the original method.The research results have good implications for object detection for fish passage in fishways.