Bird Postures Recognition Model Fusing Global and Random Local Features

In order to efficiently classify bird postures,a bird postures recognition model based on global and random local features fusion is proposed.Firstly,the global features of bird postures are extracted by using a multi-resolution fusion network.Then,a random localization module is used in the network’s shallow and deep,high-resolution features.It obtains the positions of the maximum value according to the randomly extracted feature maps,which is used to form a bounding box to cut the original image.In addition,the cropped local images are taken as input for the sub-classification network to extract the local features of bird postures.Subsequently,the global and random local features are fused,and the multi-loss strategy with the global and local losses is adopted to adjust the network.Finally,the bird postures recognition model with fusing global and random local features is constructed.The bird images with the single intact posture in the CUB200-2011 are collected and summarized to obtain the bird postures dataset,including crouching,flying,swimming,and standing postures.Experimental results based on the dataset show that the recognition accuracy of the model is better than that of the popular convolutional neural networks and achieves 96.1%.The results of ablation experiments on the random localization module and its internal randomness,grouping situation,and multi-loss strategy show that the random localization module and the multi-loss strategy can improve the recognition accuracy,which proves the effectiveness of the random localization module and the multi-loss strategy.