Research On Fine-grained Image Recognition And Classification Algorithm

Fine grained image face recognition is an important research direction in the field of computer vision,fine grained image recognition is different from the general level based on the recognition of the shape of the object,is based on the comparison of specific details,fine grained image class accuracy needs to be more detailed,through the fine local object difference effectively distinguish different instances;The difficulty of face recognition lies in the subtle visual differences of subclasses,and it is easy to be interfered by many factors.Under the condition of high visual similarity of different faces of vision,the subtle image differences are mined and the correct classification of images is carried out.Without professional training,it is difficult to do fine-grained classification,so finegrained image recognition and classification in the future face image fine processing convenient role will be immeasurable.First of all,the intra-class difference of face pictures is large,but the inter-class difference is small.Under different environment,expression,age and other interference factors,the model has significant differences in the performance of individual recognition.How to reduce the internal change,expand the external difference to obtain the feature representation of high representational power is the first main task.Secondly,too many parameters in the training process lead to low training efficiency.In most cases,the model reads static two-dimensional pictures,and a 256×256 RGB color map contains 196,608 digits.Therefore,how to speed up the convergence process of the network model is another major task.A Channel Attetion Multi-Scale Fusion residual neural network(CAMF-Res Net)based on channel attention module is proposed.The main work of this paper is as follows:(1)In view of the serious loss of recognition accuracy of top-level features extracted from mainstream classification models,and the reduction of convergence ability caused by the deepening of network is not conducive to training,residual blocks are adopted to solve the defects of increasing error rate and decreasing convergence caused by the deepening of network based on the above problems.(2)Build a multi-scale feature pyramid fusion residual network,extract feature pyramid to extract features of different levels for information fusion,and then integrate features of different scales to obtain more powerful descriptive information.Moreover,the characteristics of residual structure determine that even the deep-level network can reduce the loss of key information and greatly reduce the calculation parameters.It avoids the defect that the network training speed is seriously slow with the deepening of network depth and the final feature representation greatly improves the performance of image representation.(3)Aiming at the large number of parameters that need to be set manually when using random decline algorithm to train network models,an efficient BN layer module is proposed to solve the above problems.Usually before adding Re LU activation units to each network layer,regularization is performed on the feature maps extracted from each layer,and then nonlinear mapping operations are performed to accelerate the training convergence process of the network model.(4)In the end-to-end network feature extraction,the channel attention mechanism is used to extract the corresponding high-order feature representation,and the multi-channel attention weight is based on batch normalization through operations such as convolution pooling to obtain the image representation with strong description ability.Finally,the high-level feature representation obtained after the fusion of multi-scale feature representation and attention can be used as image classification.Experimental results show that the proposed method can effectively solve the defects of deep network degradation and excessive parameters,improve the refinement ability of the discriminant region of fine-grained images,and improve the classification accuracy of the model to obtain more descriptive information feature representation.