An Image retrieval System for Brand Logos Based on Deep Learning

摘要

#$%^&*AU2019101149A420191031.pdf#####ABSTRACT It is an image retrieval system of different kinds of sport brand logo based on deep learning, the invention is embodied through following steps: to begin with, various categories of images of sport brand logo are acquired by web crawler, which is followed by image data processing. In this step, raw data are prepossessed by rotating, flipping, adding Gaussian filtering, augmenting the contrast brightness and resizing, and then the productions are standardized processing include transforming the label to the one-hot vector and changing three-channel pictures to single channel, and these refined data are divided into training set and test set. Thirdly, the model is built which consist of 4 convolutional layers, 2 pooling layers, and 2 fully connected layers. Finally, the set of data is put into the neural network in batches and several optimizations such as adjusting the base learning rate are applied to achieve optimal performance. The average accuracy of the logo recognition fluctuates around 98.6%. Overall, this model can recognize different kinds of sport brand logo in high speed and precision without human involvement. 1