Fine-Grained Recognition Of Yunan Wild Bird Images Based On Deep Learning

Fine-grained image recognition aims to a more detailed classification of coarse-grained categories.Due to the subtle inter-class and intra-class variation between sub-categories,fine-grained image recognition is a more challenging task than the general image recognition.The thesis takes Yunnan wild bird images fine-grained recognition based on deep learning as main research topic,and focuses on the Yunnan wild bird images fine-grained classification research based on deep learning and Yunnan wild bird images fine-grained object detection research based on deep learning.Collection and labeling of Yunnan wild bird image dataset YUB-200-2017.In order to support the study of Yunnan wild bird images fine-grained recognition,the author collected 200 categories of bird images from Yunnan wild bird resources,with 60 images in each category and a total of 12000 images.At the same time,every bird image in the dataset was labeled with:(1)bird category label,(2)bounding box of bird object,(3)bounding box of bird's head,(4)bounding box of bird's body,(5)bird segmentation mask,(6)13 semantic parts of bird,(7)appearance characteristics,(8)habitat environment,(9)geographical distribution.This dataset and its rich labeling information have laid a solid foundation for the Yunnan wild bird images fine-grained recognition and related research.Yunnan wild bird images fine-grained classification based on model fine-tuning and transfer learning.Based on transfer learning and model fine-tuning,a comparative study of different convolution neural networks on the Yunnan wild bird images dataset was carried out with three modes(training from scratch,local fine-tuning and global fine-tuning).Experimental results show that global fine-tuning can obtain higher classification accuracy compared with other two strategies.For small and medium datasets,using pre-trained model on the ImageNet and global fine-tuning is the first choice to obtain better classification accuracy.Meanwhile,deep residual networks and densely connected convolutional networks have the highest classification accuracy among these convolutional neural networks.It makes the deep residual networks and densely connected convolutional networks become the first choice of the convolution neural network architectures using model fine-tuning and transfer learning.Joint semantic parts for fine-grained bird images classification.An fine-grained image classification model of joint semantic parts is proposed.The model consists of two modules,one module is a parts detector network,and another module is a three-stream classification network.By casting parts detection as three categories of object detection task,bird semantic parts can be detected accurately using the object detection algorithm based on deep residual networks.Then a three-stream deep residual networks is built to aggregate the features of object level and part level.The proposed model has higher semantic parts detection accuracy and classification accuracy on the YUB-200-2017 and CUB-200-2011.Bird images fine-grained object detection based on data augmentation and region-based fully convolutional networks.Based on the region-based fully convolutional net-works,the detection framework uses deep residual networks as backbone and consists of two stages for fine-grained bird objects detection.First,region proposal network is used to generate the candidate region bounding boxes through the residual network automat-ically.Second,given the candicate bounding boxes,position-sensitive score maps and position regression score maps are generated on the last convolutional layer of the residual networks.The detection network is learned end-to-end with multi-task loss function,and the detection results are finally obtained by non-maximum suppression.In order to enhance the robustness of the model and reduce over-fitting,data augmentation is used to expand the training set.Experimental results show that the proposed model can locate and classify multiple bird instances in the image at the same time.