Research On Fine-Grained Image Classification Method Based On Feature Fusion

Fine grained image classification is an important research field in computer vision.Different from the coarse-grained classification task,fine-grained image classification needs to further distinguish the subtle differences between the subclasses of a given object category,and the variance between the subclasses is small,so it is very challenging.The key of fine-grained image classification is to detect the local features with discriminative information among subclasses.Although the deep convolution neural network can effectively extract the basic contour and texture information of the image,it can not meet the requirements of fine-grained image classification due to the limited ability of feature expression.To solve these problems,this thesis proposes a fine-grained image classification method RPN-SCA-BCNN(RSCAB)based on weak supervised feature fusion,which is composed of target acquisition network and image classification network.The target acquisition network optimizes the regional recommendation network through the improved non maximum suppression algorithm,which can accurately locate the target object,avoid the interference of background information and reduce over fitting.The image classification network takes the processed image as the input of the network.In order to overcome the confusion of categories,a new attention mechanism Spatial-Channel Attention(SCA)is proposed,which combines spatial and channel attention in CNN to enrich the image local target features.Image classification network integrates multi-layer feature mapping to encode visual attention(spatial attention and channel attention),which improves the problem of insufficient expression ability of single scale feature.The RSCAB method proposed in this thesis does not rely on manual annotation information,only relies on the category label of the image,assigns different weights to the final output feature map,suppresses the influence of useless information on classification,and improves the classification accuracy of RSCAB method.The RSCAB method is tested and analyzed by CUB-200-2011,Stanford Cars and Oxford Flowers datasets.The experimental results show that the classification accuracy of RSCAB method can reach 86.9%,93.4% and 98.3% respectively,which is 2.8% and 2.1% higher than the original B-CNN model on CUB-200-2011 and Stanford Cars datasets.