模糊水下图像多增强与输出混合的鱼类检测方法

Fish Detection Method of Multiple Enhanced and Outputs Blend for Blurred Underwater Images

针对模糊水下图像增强后输入鱼类检测模型精度降低的问题,提出了模糊水下图像多增强与输出混合的鱼类检测方法。利用多种图像增强方法对模糊的水下图像进行增强,将增强后的图像分别输入鱼类检测模型得到多个输出,对多个输出进行混合,然后利用非极大抑制方法对混合结果进行后处理,获得最终检测结果。YOLO v3、YOLO v4 tiny和YOLO v4模型的试验结果表明,对比原始图像的检测结果,本文方法的检测精度分别提高了2.15、8.35、1.37个百分点;鱼类检测数量分别提高了15.5%、49.8%、12.7%,避免了模糊水下图像增强后输入鱼类检测模型出现精度降低的问题,提高了模型检测能力。

The underwater images of aquaculture ponds,rivers and sea inlets were generally fuzzy and low contrast due to the influence of water turbidity and light attenuation in water.However,the existing literature found that the clarity brought by image enhancement cannot directly improve the detection ability of fish detection model,and even the detection accuracy of the model was degraded.An multiple and outputs blend enhanced method was proposed for fish detection.Blurred underwater images were enhanced by various image enhancement methods,and the enhanced images were input into the fish detection model to obtain multiple outputs.Then the mixed results were postprocessed by non-maximal inhibition method to obtain final test results.Compared with the detection results of the original image,the experimental results on YOLO v3,YOLO v4 tiny and YOLO v4 models showed that the detection accuracy of the proposed method was improved by 2.15 percentage points,8.35 percentage points and 1.37 percentage points,and the number of fish was increased by 15.5%,49.8%and 12.7%,respectively.The proposed method achieved the purpose of improving the model detection ability,and it can be applied to fish count and fish category detection.