Research On Flower Image Classification Based On Deep Learning

Flowers and plants are an indispensable part of our lives.In the process of protecting flowers and plants,if only humans are used to identify and classify flowers,it will cost a lot of manpower and material resources.How to use the machine to automatically classify flowers accurately is a hot issue that needs to be solved at present.Accurate classification of flower images is also a necessary prerequisite for flower plant management to move towards artificial intelligence.Deep neural networks have achieved good classification results in coarse-grained image classification,but for fine image classification such as flower images,the effects of deep neural networks in flower image classification are not yet ideal.Because the flower image belongs to the sub-category under the big flower category,there are characteristics of large intra-class differences and small inter-class differences,and the classification task is more difficult.This article takes flower images as the research object,the main research contents are as follows:(1)Based on the Xception convolutional neural network,this paper proposes a multi-loss spatial attention network,a multi-loss channel attention network,and a multi-loss multi-attention network.Conclusion Xception,spatial attention and channel attention design spatial attention residual module,channel attention residual module and multi-attention residual module.These three modules were added to Xception to strengthen the positioning and feature extraction capabilities of the flower image area;at the same time,the network's loss layer combines triplet loss(triplet loss)and classification loss(softmax loss),so that the network has a simultaneous The features of higher intra-class compactness and inter-class separation are embedded in the space,thereby improving the classification accuracy of flower images.Experiments on the two flower images Oxford 17 flowers and Oxford 102 flowers data sets show that the three network models proposed in this paper have greatly improved the accuracy of flower image classification.(2)The two most important tasks in the process of flower image classification are the precise positioning of the flower image area and the feature extraction of the positioning area.For these two tasks,two networks are used to complete the two tasks respectively,thereby further proposing multiple losses Attention bilinear network classifies flower images.Taking ResNet50 and Xception with stronger feature extraction capabilities as the basic network,the spatial attention residual error module and the channel attention residual error module are added to the two networks respectively,which are responsible for the location and feature extraction of the flower image area.The two networks coordinate with each other to complete the classification of flower images,thereby further improving the classification accuracy of flower images.The experiments on the two flower images Oxford 17 flowers and Oxford 102 flowers data sets show that the classification accuracy of the multi-attention attention bilinear network proposed in this paper on flower images has been further improved.