Research On Multi-scale Detection Algorithm Of Invasive Pulmonary Tuberculosis Based On Attentional Mechanism

With the continuous development of deep learning technology,the convolutional neural network has shown better performance under various complex visual challenges,which makes it possible to establish a set of computational auxiliary diagnostic technology based on deep learning.Compared with manual diagnosis.The use of computer aided technology has the characteristics of low cost and high efficiency.However,in the medical field,the computer diagnosis technology often has a very high requirement on the accuracy of the model.The classification of medical images can be roughly divided into 2d image classification and 3d image classification.There have been many different studies on the classification of 2 dimensional images.In order to improve the fitting ability of the network,most of the multi-branch structures are used,either in the form of single branch network and data set with different resolutions,or in the form of multi-branch network and single data set.As a result,the network structure is bloated and the network training speed is too slow.Secondly,the traditional convolutional neural network mostly uses the output feature information of the last layer to classify,and ignores the utilization of low-level semantic information of the network.In this paper,aiming at the problem of small lesion area and indistinct positive and negative sample differentiation in medical images,an progressive multi-scale information fusion mechanism is designed.Combining with the attention mechanism in natural image classification,the existing convolutional neural network structure is improved,and the end-to-end automatic image recognition model is established.The main work of this paper is as follows:The cross-layer connection structure in residual neural network is used to construct the network.The causes of network degradation in convolutional neural network are analyzed.The use of cross-layer connection can avoid the occurrence of network degradation,which makes it possible to build a convolutional neural network model with a large number of layers.The increase in the number of parameters and nonlinear transformation is conducive to the improvement of network generalization ability and the accuracy of the model.At the same time,by referring to the common optimization methods in convolutional neural network,the convergence speed of the network is improved by means of batch normalization and 1*1 convolution.The model can be improved by the attention mechanism based on human vision system.In attention mechanism,first of all,the original characteristics of the semantic information is compressed inside the figure,based on the feedforward neural network structure characteristics of the original figure the relationship between different channel operation,operation is used to derive the weight vector of the characteristics of the original figure,new features in different proportion of local characteristics can be changed,the network of key areas of focus for promotion,and attention mechanism have not change the characteristics of the network input dimension,can be flexible in convolution neural network applications.Some solutions are proposed to solve the problems of semantic information utilization,the effectiveness of fusion strategy,and the information loss in the lower sampling.Firstly,the semantic information of different layers of the network is fused by the method of extended cross-layer connection,which increases the utilization of low-level information of the network.Then,the network hierarchy is divided,and different fusion schemes are designed for different layers of the network.Finally,the maximum pooling is used in the next sampling process to select more meaningful elements in the pooling area.The final multi-scale information fusion mechanism is designed based on three problemsIn this paper,we set up different comparisons based on the number of layers,whether to use attention mechanism,whether to use multi-scale feature fusion mechanism and whether to use both at the same time.The performance of the network as a whole is described with different evaluation indexes of precision,precision,recall and F1 value.In the optimal case,the network precision can reach 94.45%.Experimental results show that the improved model is superior in time and space complexity,and that the attention mechanism and multi-scale information fusion mechanism used in this paper can improve the accuracy of image recognition.At the same time,some biased data in the experimental results are analyzed,and the causes of these data are explored by combining with the structure of the network.