基于斑马鱼图像特征的鱼群检测算法北大核心CSCD

计算机视觉方法越来越多地应用于斑马鱼的群体行为研究;但是,由于斑马鱼游动过程形体变化大,遮挡多,准确与鲁棒地检测出斑马鱼仍然是一件非常具有挑战性的问题。为了解决该问题,提出一种基于斑马鱼图像特征的鱼群检测算法。首先通过分析目标特性,提出使用鱼头和鱼尾替代全鱼的检测方法,解决了传统整鱼检测在鱼群交叉遮挡时失效的难题;然后基于斑马鱼图像特征自动构建训练集,避免了深度学习手动标注的费时费力问题。通过对实际斑马鱼视频进行处理验证,与现有的算法相比,本文提出的方法在标注率、召回率(recall,R)与遮挡检测率(occlusion detection rate,ODR)等性能指标上有更好的实验效果。其中,在标注性能方面,本文提出的自动标注方法在总标注率上达到87.40%;在训练集效果方面,本文自动标注算法结合人工校正在标注时间上相比于人工标注方法减少93.11%,均值平均精度(mean average precision,mAP)达到79.80%;在目标检测方面,在目标遮挡率为42.72%的情况下,本文检测算法能够获得82.0%的召回率及58.02%的遮挡检测率。

Fish stock detection algorithm based on zebrafish image features

The computer vision methods are increasingly applied to the study of zebrafish stock behavior. However, due to the large form changes and many occlusions in the swimming process of zebrafish, it is still a very challenging problem to accurately and robustly detect the zebrafish. To solve this problem, a fish stock detection algorithm based on zebrafish image features was proposed. Firstly, by analyzing the target characteristics, a detection method using fish head and fish tail instead of the whole fish was proposed, which solved the problem that the traditional whole fish detection failed in the case of fish stock cross occlusion. Then, the training set was automatically constructed based on zebrafish image features, which avoided the time-consuming and laborious problem of manual annotation in deep learning. Through the processing and verification of the actual zebrafish videos, and compared with the existing algorithms, the proposed method has better experimental effects in the performance indexes such as annotation rate, recall rate and occlusion detection rate (ODR). In terms of annotation performance, the total annotation rate of the proposed automatic annotation method is 87.40%. In terms of the effect of training set, the automatic annotation algorithm combined with manual correction reduces the annotation time by 93.11% compared with the manual annotation method, and the mean average precision (mAP) is 79.80%. In terms of target detection, when the target occlusion rate is 42.72%, the proposed detection algorithm can obtain the recall rate of 82.0% and the occlusion detection rate of 58.02%.