Research On Fine-grained Image Classification Based On Deep Convolutional Neural Network

As one of the most basic problems in the history of computer vision,image recognition has been widely studied.in recent years.Due to the continuous increase in the scale of data,the research goals have also changed dramatically.This paper studies one of the latest and most challenging object recognition tasks,Fine-Grained Visual Classification(FGVC).The problem of fine-grained image classification is to classify sub-categories.The difference and difficulty of FGVC compared to general image recognition task represented by Image Net image classification is that the granularity of the category to which the image belongs is more fine.Aiming at the problem of fine-grained image classification,this paper analyzes the current research status at home and abroad.Considering the fine-grained image recognition with strong supervision information have made a great progress,but it requires a lot of manual annotation information,consumes a lot of manpower and material resources,and is not conducive to actual needs.Starting from fine-grained image recognition with weak supervision information to improve the performance of fine-grained image recognition,the main work of the paper includes:1)At present,the Res Net and Inception networks are widely used in the field of image recognition,and have achieved excellent performance on the Image Net dataset.A network model combining the two networks is designed to improve the performance of fine-grained image recognition.2)The activation function commonly used in convolutional neural networks is the Rectified Linear Unit(Re LU).In this research process,PRe LU was used as the activation function instead of Re LU in the above residual connection.PRe LU is an improved version of Re LU,adding a learnable parameter so that the slope of the negative part can be determined based on the data.3)Considering that many innovative convolutional neural networks have appeared recently,and they have also achieved brilliant results in the field of image classification.For the Bilinear CNN model,use the better-performing Inception network as a feature extractor instead of the original feature extractor to improve the performance of the Bilinear CNN,thereby improving the Bilinear CNN for finegrained images recognition ability.The experimental results on CUB200-2011 and Stanford Cars fine-grained image datasets show that the designed single network model has better classification performance than the single Res Net and Inception network,and the improved Bilinear model achieves 88.3% and 94.2% classification accuracy on the two data sets,respectively,which is comparable to some previous strong supervision algorithms,which verifies that the validity of our method.