【摘要】 【目的】借助计算机视觉技术检测鱼群的摄食行为变化,为实现水产养殖中的精准投喂提供参考。【方法】以彩鲤为试验对象,首先对鱼群摄食过程的图像进行中值滤波、直方图均衡化预处理,提取目标鱼群前景图像,然后通过灰度共生矩阵计算鱼群的纹理特征,利用Lucas-Kanade光流法计算鱼群的运动方向矢量,并采用方向熵表征鱼群运动的混乱程度,最后利用支持向量机算法对鱼群摄食图像进行训练检测。【结果】图像预处理方法减少了水质浑浊对鱼群图像检测的影响;基于光流法与图像纹理特征定量分析鱼群的摄食过程变化,避免了水面抖动及水花等复杂因素的影响。该方法的检测准确率达97.0%,基本上能检测识别出所有的鱼群摄食状态与正常游动状态图像;与基于形状和纹理特征的检测方法相比,其检测精度提高4.5%(绝对值),可更好地满足池塘养殖环境下的鱼群摄食行为检测工作需要。【建议】在今后的研究中将获取更多鱼群在不同环境下的摄食图像,以提高模型泛化性,同时对鱼群不同摄食阶段特征参数设定阈值告警,以便更好地应用于智能饵料投喂设备研究。更多还原

【Abstract】 【Objective】Using computer vision technology to detect changes in feeding behavior of shoal could provide a reference for accurate feeding in aquaculture.【Method】Taking Cyprinus carpio var. color as an experimental object,median filtering and histogram equalization pre-processing were used to extract the foreground image of the target shoal,and the texture features of the shoal were calculated by the gray level co-occurrence matrix. The direction vector of the shoal was calculated by Lucas-Kanade optical flow method,then used direction entropy to characterize the chaos of shoal movement. Finally,the support vector machine algorithm was used to train and test the shoal feeding image.【Result】In this paper,the image preprocessing method was used to reduce the influence of water turbidity on fish image detection.The optical flow method and image texture features quantitatively analyzed the changes in the feeding process of the shoal,avoiding the influence of complex factors such as water surface jitter and water splash. The experimental results showed that the detection accuracy of this method reached 97.0%,and it could basically detect the images of all the shoal feeding state and normal swimming state. Compared with the detection methods based on shape and texture features,the detection accuracy was improved by 4.5%(absolute value),this method could better solve the problem of shoal feeding behavior detection in pond culture environment.【Suggestion】In the subsequent study,more shoal feeding image in different environments will be used to improve the generalization of the model. In addition,threshold alarms will be set for the features parameters of different feeding stages of the fish to better apply to the research of intelligent bait feeding equipment.更多还原