Research On Bird Species Recognition Based On Deep Learning

With the deterioration of global ecological situation,birds have been more threatened than ever before.It is urgent to develop automatic bird recognition techniques for protecting birds and conserving the species diversity in the Earth.The computer vision-based bird species recognition is usually considered as a fine-grained visual categorization(FGVC)problem whose target is to distinguish subordinate-level categories.Among all the FGVC problems,bird species recognition is one of the most complex tasks due to its high intra-class variance and low inter-class variance.The difficulties of bird species recognition are listed below.First,there usually is complex background noise in most of the bird images while subtle variance among different categories often lies in some specific key parts.Second,simple features are not discriminative enough for bird recognition due to the complex structure of the images caused by changeable pose,view,and illumination condition.Third,due to the difficulties of collection and annotation,the amount of bird image data is limited,which raises the risk of overfitting.To this end,we engage in bird species recognition based on deep-learning and propose a recognition system based on the part-based idea.Four modules are applied in the proposed algorithm including object and part detection,data augmentation,feature extraction and classification.Single Shot Multibox Detector is adopted in the object and part detection module to locate the bird object and key parts in the image.The target of object detection is to retrieve the foreground object and filter out the background noise.And the part detection is to localize key regions which are important for the categorization task.The proposed recognition system obtains a multi-branch structure from this module.Then the data augmentation module applies Deep Convolutional Generative Adversarial Networks to augment image data of birds which can reduce the risk of overfitting.The idea of Gaussian-based second-order pooling is applied in the feature extraction module,including Robust Estimation of Approximate Infinite Dimensional Gaussian and Matrix Power Normalized Covariance.High order information has a better ability to express image information.Finally,the classification module first chooses proper machine learning algorithms for categorization and then uses the idea of ensemble learning to achieve the multi-branch fusion task.Averaging and learning strategy of decision fusion are adopted in this module respectively.A unified recognition system can be obtained from the ensemble learning-based classification module.The proposed bird recognition system based on the above modules can elegantly solve the difficulties of bird recognition and can obtain a good performance on the recognition task.Multiple experiments are conducted on the standard bird dataset CUB200-2011 to verify the classification ability of the proposed system and to evaluate the effectiveness of different modules.With no annotations in the testing phase,the proposed recognition system achieves 89.7% classification accuracy on the validation set of CUB200-2011,which is superior to most of the state-of-the-art methods.