基于DCNN特征融合的细粒度病鸡识别研究

Research on Recognition of Fine-grained Sick Chicken Based on DCNN Feature Fusion

病鸡区别于健康鸡的视觉特征主要表现在鸡头复合特征、鸡身纹理特征和体态特征.本文提出区域深度特征融合的细粒度病鸡识别方法,模型通过区域定位和分类识别两个阶段实现,区域定位采用Faster R-CNN模型对鸡、鸡头和鸡身进行检测,分类网络通过融合有区分度的语义区域特征,应用CNN网络进行病鸡分类识别.实验采用自制的数据集,通过迁移学习的方法,在共享特征图的基础上,进行分阶段的模型训练和分类.模型在语义区域检测实验中,准确率达到了99%,召回率达到了95%左右,表征模型效果的综合评价指标F1值达到了97%左右,病鸡识别中融合特征的分类精度达到了82.66%.实验结果表明:本文方法对病鸡识别有较好的效果,为在养鸡场实际场景中进行病鸡的实时识别和鸡情监控提供了技术支持.

The visual characteristics of sick chickens different from healthy chickens are mainly manifested in the compound characteristics of chicken heads,the texture characteristics of the chicken body and the characteristics of body posture.A fine-grained sick chicken recognition method based on regional deep feature fusion is proposed in this paper.The model is implemented by two-stage network of regional positioning and classification.Faster R-CNN model is used by the regional positioning to detect chicken,chicken head and chicken body.Based on the characteristics of the semantic region,the CNN network was used to classify the sick chickens.A self-made data set,and the transfer learning method to perform staged model training and classification on the basis of shared feature maps are used in the experiment.In the semantic region detection experiment of the model,precision reached 99%,the recall rate reached about 95%,and the comprehensive evaluation index F1 value representing the model effect reached about 97%.The classification accuracy of the fusion features in the recognition of sick chickens reached 82.66%.The experimental results show that the proposed method has a good effect on the recognition of sick chickens,and provides technical support for real-time recognition of diseased chickens and monitoring of the situation of chickens in the actual scene of chicken farms.