基于改进YOLO-MAO检测框架的笼养白羽肉鸡行为检测方法OA北大核心CSTPCD
Behavior Detection Algorithm for Caged White-feather Broilers Based on Improved YOLO Detection Framework
在大规模的肉鸡养殖场中,肉鸡行为通常由饲养员或专业兽医观察和分析,以确定肉鸡健康状况和养殖环境状态.然而这种方法耗时且主观.此外,在笼养环境中,由于鸡只的高密度和严重的互相遮挡,行为的视觉特征不明显,传统的检测算法不能准确地识别鸡只的行为特征.因此,本文提出一种改进的笼养白羽肉鸡行为检测的目标检测算法.所提出的算法由2个模块组成:多尺度细节特征融合模块(MDF)和目标关系推理模块(ORI).多尺度细节特征模块充分利用和提取网络浅层特征映射中包含的多尺度细节特征,并将它们融合到负责相应尺度检测的特征映射中,实现细节特征的有效传输和补充.目标关系推理模块充分利用对象之间的位置关系进行推理和判断,使模型能更充分地利用对象之间的潜在关系来辅助检测.为验证所提出算法的有效性,在目标检测领域具有权威性的COCO公共数据集以及真实的大规模笼养白羽肉鸡养殖环境中自建的行为检测数据集上进行大量对比实验.实验结果表明,与其他最先进的模型相比,本文所提出的改进算法在COCO数据集和自建数据集上均达到最佳识别准确率;对喂食、饮水、移动和张嘴等影响肉鸡健康状况较为重要的行为进行检测,识别精度分别达99.6％、98.7％、99.2％和 98.3％.
In large-scale broiler farms,the behavior of broilers is usually observed and analyzed by feeders or professional veterinarians to determine their health status and breeding environment status.However,this method is time-consuming and subjective.In addition,in caged environments,due to the high density of chickens and serious mutual occlusion,the visual features of behavior are not obvious,and traditional detection algorithms cannot accurately identify the behavior characteristics of chickens.Therefore,an improved object detection algorithm for behavior detection of caged white-feather broilers was proposed.The proposed algorithm consisted of two modules:multi-scale detail feature fusion module(MDF)and object relation inference module(ORI).The multi-scale detail feature module fully utilized and extracted the multi-scale detail features contained in the shallow feature maps of the feature extraction network,and integrated them into the corresponding feature maps responsible for detection at the corresponding scale,achieving effective transmission and supplementation of detail features.The relational reasoning module fully utilized the positional relationships between objects for inference and judgment,enabling the model to more fully utilize the potential relationships between objects to assist in detection.To verify the effectiveness of the proposed algorithm,a large number of comparative experiments on both authoritative public datasets in the field of object detection and self-built behavior detection datasets in real large-scale caged white-feather broiler breeding environments was conducted.The experimental results showed that the proposed improved algorithm achieved the best detection accuracy compared with other state-of-the-art models,both in the COCO dataset and the self-built dataset.For the detection of behaviors such as feeding,drinking,moving,and opening the mouth,which were crucial for the health status of broiler chickens,the algorithm achieved accuracy rates of 99.6％,98.7％,99.2％,and 98.3％respectively.
夏元天;寇旭鹏;薛洪成;李林
中国农业大学信息与电气工程学院,北京 100083
计算机与自动化
白羽肉鸡行为识别目标检测多尺度细节特征融合模块关系推理模块
white-feather broilersbehavior recognitionobject detectionmulti-scale detail feature fusion modulerelation inference module
《农业机械学报》 2024 (011)
103-111 / 9
国家科技创新2030—新一代人工智能重大项目(2021ZD0113701)
评论