Bilinear pyramid network flower image classification method

Abstract

The invention discloses a bilinear pyramid network flower image classification method, and the method comprises the steps: adjusting the size of a to-be-classified flower original image to 224*224, and randomly cutting the to-be-classified flower original image to 192*192; performing feature extraction on the flower image through a bilinear pyramid network according to the adjusted image; and finally, inputting the extracted features into a classifier for classification, and outputting the classified features to obtain a classification result of the flowers. According to the method, a flower classification bilinear pyramid network is adopted, and features of a convolutional layer and features of the convolutional layer are fused instead of directly inputting a final classifier through the network. The features coded from the feature pyramid automatically carry multi-level semantic clues, have additional robustness for changes of postures and scales, and are superior to single-layer features in the aspect of classification and recognition. Wide verification is performed on the reference data set to display the validity of the proposed method.

Description

technical field [0001] The invention relates to the technical field of fine-grained image classification, in particular to a bilinear pyramid network flower image classification method. Background technique [0002] Fine-grained visual classification refers to distinguishing like species from the same domain (e.g. identifying fine-grained breeds of dogs or specific car models). Although fine-grained species usually look similar to each other, different images in the same domain may have different poses, positions, etc., making them different, so larger intra-class differences and smaller inter-class differences give fine-grained Granularity species classification poses serious challenges. [0003] The most similar prior art implementation of the present invention: [0004] Research in cognitive science has shown that basic-level categories are distinguished by differences in body parts, while fine-grained species are distinguished by different attributes of the same parts....

AI Technical Summary

Problems solved by technology

[0006] First, handcrafted features cannot distinguish subtle differences between visually similar species; second, feature extraction and final classification are performed independently, which hinders the global optimization of classification models; third, only high-level visual cues, while ignoring the importance of underlying visual cues

Images