Study On Distracted Driving Image Classification Algorithm Based On Bilinear Attention Network

With the rapid development of automobile industry,traffic pressure is becoming more and more serious,and traffic safety has become the focus of people’s attention.The detection of distracted driving behavior has become an active field in image classification research.How to accurately and efficiently detect the distracted driving behavior with insignificant overall dynamics,small motion variation range and small action space in the scene inside the car has become the focus of distracted driving image detection.At present,traditional CV algorithm or deep learning method is often used for distracted driving image classification,but these methods all have the following problems :(1)Traditional research methods based on computer vision have high requirements on the environment,narrow application scope,multiple parameters and large computation amount,and complicated calculation method.(2)Existing image classification models for distracted driving are not ideal in generalization ability,and the problem of difficult samples is not well solved.There is always a small part of images that are misclassified due to the lack of obvious classification features.(3)Due to the single background environment of driving images,invalid information exists in most areas of the images,resulting in redundancy of irrelevant image information and poor impact on classification accuracy.To solve the above problems,this paper takes the data set of distracted driving published on the Internet as the research object,and combines the bilinear pooling model,the difficult sample processing method and the improved dual-path attention mechanism to conduct information mining to improve the classification accuracy.The work of this paper is as follows:(1)in view of the existing classification of distracted driving model parameter,the computation is large,complex calculation method,build the new compact double linear convolution neural network,the network can extract has the specificity of distracted driving behavior characteristics at the same time,effectively avoid the high cost,computation and large storage capacity,and the situation of the gradient to disappear.The compact bilinear pooling model proposed in this paper can effectively reduce the problem of distracted driving image classification error.(2)In view of the problem that the generalization ability of distracted driving image classification model is not ideal and there are always some image classification errors in training,this paper carries out the difficult sample mining and proposes the multiple loss network structure,so that the model can focus more on the difficult sample classification during training.(3)Aiming at the problems of distracted driving image information redundancy,lack of specificity and poor classification accuracy,a double-path convolutional neural network with attention mechanism was proposed to effectively eliminate useless features and reduce redundancy.In this paper,the performance of Distracted driving image recognition is analyzed by State Farm Driver Detection,an open data set.The confusion matrix is used as the evaluation criterion,and the results show that the proposed bilinear neural network algorithm is more accurate,which proves that the bilinear convolutional network model presented in this paper has better image classification effect.