Fine-Grained Image Recognition Based On Clicked Deep Model

As a popular research direction in recent years,fine-grained image recognition plays an important role in various fields in the industry.However,since its goal is to distinguish the differences between the subordinate,the target of the task is visually small differences lead to a very large challenge to the fine-grained image recognition task itself.At present,one of the-state-of-art in fine-grained image recognition is the bilinear convolution neural network,whose main idea is to recognize the images with rich visual features.Although good results have been achieved,the performance still remains insufficient.In order to improve the recognition accuracy,this paper constructs a click feature that can be used to express the semantic information of images by introducing click data to solve the problem of semantic gap.Based on that,we further proposed the construction method of click depth model.This new model is called C-BCNN(Click guided Bilinear Convolutional Neural Network)by combining the user's click data to capture visual data and semantic content in the image,then using visual feature and the click feature to identify the image.Noted,large-scale click data also poses serious noise and data imbalances.In response to this,we propose a weak supervised learning method to train the model parameters.This model,which combines weakly supervised learning methods,is called W-C-BCNN.This new model automates the weighting of samples by estimating the degree of picture reliability,so that more reliable samples contribute more to training The results show that: 1)Click features based on click data do have some effect in fine-grained image recognition;2)By integrating the visual features extracted by the bilinear convolutional neural network and the click features constructed by the click data effectively improve the effect of fine-grained image recognition.3)Click data and visual consistency can effectively help to model the degree of image reliability.From the results of the extensive experimental data,the model method incorporating the click feature is 18% lower than the traditional bilinear convolution neural network in training error,while the model recognition effect of joining the weak supervised learning is more than the one depth model C-BCNN,which further increased by 2%.Thus,our experimental method does improve the fine-grained image recognition.