Research On Image Processing Based On Recurrent Attention Model

In computer vision tasks,deep learning technology makes the machine recognition accuracy has been better than human,so more and more scholars focused on the model performance bottleneck,and proposed paper and models.The main reason why the classical model is generally computationally expensive is that the model is essentially a window-recognition model,which is based on the computational performance of the machine.Aiming at the problem of high computational cost,this paper puts forward a new method based on Partially Observable Markov Decision Processes model.The model extracts the local part from the input image and performs the operation.The model has many advantages,including: the total amount of parameters and the amount of computation independent of the size of the input image,strong robustness,end to end training.But the model still has great room for improvement: the focus is relatively small,excessive cycle times of extraction;for natural scene image recognition rate is low;to be used in computer vision tasks important target recognition etc..To solve the above problems,this paper puts forward a multi focus focusing cycle model,reduce the times of extraction model cycle,improve the recognition speed;Recurrent attention model based on convolution,take the advantages of convolutional network over recognition of natural scene image data;Target detection model based on recurrent attention model,appling recurrent attention model to identify the target in the input image,including categories,borders,and coordinates.Specific research work is as follows: 1,a multi attention model is proposed.The recurrent attention model concentrates the computational resources on the data specific region.Compare with the input image and the identification of the target,the focus area is usually smaller;if the focus area is too small,it will lead to excessive number of iterations,reducing efficiency,it is difficult to find multiple targets in the same input.In this paper,a multi attention model is proposed,which can be used for multiple parallel focusing.The use of reinforcement learning training,all focus on the behavior of uniform scoring training.Compared with the single focus model,the training speed and recognition speed are improved.At the same time,the model has high universality.2)a recurrent attention model based on convolution is proposed.In the image recognition task,recurrent attention model can effectively identify the true number or the number of images,such as Google Street brand,but can not effectively identify the natural scene images,such as object recognition for pervasive data sets Cifar10.To deal with the natural scene image data set,first analyzes the advantages and shortcomings of the classical recurrent attention model,combined with the characteristics of recurrent attenion model,this paper puts forward three strategies to improve the recurrent attention model based on circular convolution network,combining the advantages of convolutional network and cyclic focusing model,which can effectively identify the natural scene images.In this paper,the improved cyclic focusing model can effectively identify Cifar10 data sets.3,a target detection model based on attention is proposed.The classic model of the target detection is to scan the whole image of the input,and extract the candidate region to identify,with large amount of calculation.In this paper,a target detection model based on attention is proposed.After obtaining the input,the model dynamically extracts the images and calculates to get the information about the object and the coordinates.This model has the advantages of the end to end training mode,the independence of computation and the input size,robustness and so on.Experiments show that the model can effectively accomplish the task of target detection,and the advantage is more obvious in the larger image.