The Research On Fast Auxiliary Diagnosis Algorithm Of Snake Bite Based On Attention Mechanism

In the field of fine-grained image classification,bilinear convolutional neural network(B-CNN)has excellent classification accuracy,which is at least 10% higher than that of single linear convolutional neural network.However,bilinear convolutional neural network also has its disadvantages.The first disadvantage is that the effective information and invalid information in the image are extracted with the same importance,which makes its classification accuracy unable to be improved and prolongs the experimental time;The second disadvantage is that the model fails to fully extract image features during training,resulting in insufficient convergence of the model.In view of the above shortcomings of bilinear convolutional neural network,a model integrating attention mechanism and progressive learning method into bilinear convolutional neural network is proposed,which is combined with the practical application of rapid auxiliary diagnosis of poisonous snake bite.Firstly,aiming at the problem that the effective information and invalid information in the picture are extracted with the same importance,a method of integrating the attention mechanism with bilinear convolution neural network is proposed.In this paper,Grad-CAM is selected as the attention module of convolution neural network,and Res Net-50 is selected as the feature extractor of convolution neural network model.The purpose is to maintain the consistency of the whole model and use the same weight file,There is no need to repeat the model to save experimental time.In the improved bilinear convolution neural network model,Grad-CAM attention module needs to identify the important and non important regions in the picture,then send the two different regions into the bilinear convolution neural network model for feature fusion,and then obtain the classification results according to the contribution mechanism.In this paper,the improved bilinear convolution neural network model is trained on four different data sets.The experimental results show that the improved bilinear convolution neural network model has higher classification accuracy than the original model.Secondly,aiming at the problem of insufficient model feature information extraction,a training method using progressive learning method in the training stage is proposed.When the model extracts feature information,the model will directly learn the high-level global feature information,while ignoring the low-level local feature information in the learning process.Local feature information is also helpful for the improvement of classification accuracy.The introduction of progressive learning method forces the model to learn low-level local feature information first,and then let the model learn high-level global feature information by adding layers,so that the model can fully learn local and global features.Finally,the improved model is used to test the practical application effect of snake bite auxiliary diagnosis.Because the common poisonous snakes in China lack public data sets,the authors need to make their own data sets and collect 5000 pictures through the network,which are divided into five categories.The data set of common poisonous snakes in China is applied to the progressive attention bilinear convolution neural network model(PAB-CNN),and the classification accuracy is 84.9%.