Research On Attention Model In Image Classification And Image-to-Image Translation

Most of the outstanding results in the field of image processing at this stage are due to the continuous deepening of the number of layers in the neural network.However,as the neural network becomes more and more complex,and the parameters that the neural network needs to store more and more,it will cause network capacity problems.For image classification,it is still a challenge to improve classification accuracy without introducing too many parameters.In the field of image-to-image translation,the combination of generative adversarial networks can realize the translation of images from the source domain to the target domain.However,the transformation model using generative adversarial networks usually produces semantic artifacts,and the irrelevant part of the generated image changes greatly,which makes the generated image of lower quality.In order to solve the above problems,this paper proposes a new type of attention module—Pixel-wise And Channel-wise Attention,which is referred to as PC attention module for short.By using a variety of attention modules in the classification model,the superiority of the PC attention module over other attention modules in classification tasks is verified on the Cifar10 dataset.Moreover,it is also applied to the GoogleNet network architecture and achieved good classification performance on the Cifar100 dataset.Aiming at the special network structure of encoder-decoder used in the field of image-toimage translation,based on the PC attention model,a skip connection structure(SPC attention model)during downsampling and upsampling is proposed.During the translation process,the attention map obtained by downsampling is passed to upsampling to generate a more reasonable image.This module can help the algorithm identify the most distinctive semantic objects in the image(that is,the objects to be transformed),and can be directly applied to the convolutional neural network as a layer in the network structure to be trained end-to-end.In addition,the application of the attention model proposed in this paper to image-to-image translation is to minimize the changes of unrelated parts of the image without using additional data and networks.In the experimental part,this paper applies the SPC attention module to Pix2 Pix algorithm and CycleGAN algorithm,and compares it qualitatively and quantitatively with the latest image-to-image translation methods.It is verified that the attention model can establish a more realistic mapping from the source domain to the target domain in the image-to-image translation algorithm,thereby improving the quality of the images generated by the conversion model.At the same time,the SPC attention model is applied to the image reconstruction task,and the image reconstruction task is realized in the self-encoder algorithm using the MNIST data set.Compared with the application of other attention models in the self-encoder algorithm,this model can reconstruct more realistic images.This experiment also verifies the rationality of the PC attention model in the encoder-decoder network using the jump connection structure.