Fine-grained bird recognition method based on cross-layer simplified bilinear network

Abstract

The invention discloses a fine-grained bird recognition method based on a cross-layer simplified bilinear network. The method comprises the following steps: firstly, preprocessing 5994 training pictures and 5794 test pictures in a CUB-200-2011 data set, and then inputting the processed images into a VGG-16 convolutional neural network to extract a feature map of a bird image; in order to considerinterlayer feature interaction, three groups of simplified bilinear feature representations are extracted from obtained feature maps of different high-level convolution, and after normalization operation is performed on the feature maps, the feature maps are sent to a softmax classifier in a cascaded mode. And finally, optimizing the whole network by using cross entropy loss and paired confusion loss. The recognition method described by the invention has the advantages of low feature dimension, small calculation amount, high recognition rate, strong robustness and the like, has a certain use value for the specific field of fine-grained image classification, and can be practically applied.

Description

technical field [0001] The invention designs a fine-grained bird recognition method based on a cross-layer simplified bilinear network, which involves deep learning and fine-grained image classification. Background technique [0002] The main purpose of fine-grained classification is to distinguish its various subcategories under the same basic category, such as different kinds of birds, flowers, etc. Compared with coarse-grained images, fine-grained images have slight differences between classes and significant intra-class differences. The acquisition of fine-grained features is often more complicated and relies more on image annotations to determine complex parameters in the model, avoiding errors caused by a small amount of data as much as possible. overfitting phenomenon. The early fine-grained recognition methods rely on the local information of manual annotation to perform strong supervised learning on the classification model. Local annotation usually requires exper...

Images