Clothing Image Classification And Denoising Research Based On Convolutional Neural Network

With the development of the economy and the improvement of productivity,people’s demand for clothing is also increasing.Nowadays,people can easily buy the clothing they want through the Internet.In order to accurately and clearly show the clothing images that customers want,the research on clothing image classification and denoising has attracted extensive attention.Compared with general images,clothing images have the characteristics of rich features and unobvious differences in texture structure,so that the pooling method commonly used in convolutional neural networks cannot extract all the features in the image well,leading to the classification of clothing images is not accurate enough.At the same time,it is difficult for traditional image denoising methods to ensure the integrity of clothing image details.In view of the challenges of clothing image classification and denoising,this paper uses a deep learning algorithm based on convolutional neural networks for research.The specific research content is as follows:1)Proposed a new type of pooling method—adaptive weighted pooling.Based on feature extraction,various pooling methods in convolutional neural networks are analyzed,and it is concluded that the existing pooling methods cannot accurately extract various types of features.At the same time,the characteristics of ideal pooling methods are summarized,and a new pooling method named adaptive weighted pooling is proposed.2)Use adaptive weighted pooling for convolutional neural network to classify clothing images.For different clothing image data sets,different CNN network structures are designed,and the traditional pooling method in the network is changed to adaptive weighted pooling.In view of the problem that maximum pooling and average pooling strategies are single,they are changed to adaptive weighted pooling with the same window size.In view of the problem that stride convolution can only sample at fixed intervals,the stride convolution of 3×3 windows is changed to dense convolution of 3×3 windows followed by adaptive weighted pooling of 2×2windows.1 × 1 window stride convolution is changed to 2 × 2 window adaptive weighted pooling.In view of the problem that global average pooling does not distinguish between different features,it is changed to adaptive weighted pooling with the window size equal to the size of the feature map.The CNN network before and after the improvement is used for clothing image classification,and the results are compared.The experimental results show that the improved network structure improves the classification accuracy.3)Verify the feasibility of image denoising with structured image prior network combined with ASM energy.Experiments have confirmed that in the process of generating images by the structural image prior network,the lower the ASM energy of the target generated image,the more iterations are required to generate the image.For the same network structure,it is necessary to generate a noise image than to generate Natural images require more iterations.If an image containing noise is used as the target of the network to generate an image,during the network iteration process,the noise in the image will be generated at the later stage of the iteration.During the iteration process,a denoise image will be generated.The denoising effect of the generated image is positively correlated with its ASM energy.4)Build a UNet structure image priori network for denoising clothing image.Based on the relationship between structure image prior network image denoising and ASM energy,a 10-layer UNet network is designed,with random vectors as input and clothing image containing noise as target output.The training is stopped when the network is iterated to the maximum value of the ASM energy,and the network output at this time is the denoised clothing image.Use this algorithm and traditional image denoising algorithm to carry out clothing image denoising comparative experiment.The experimental results show that this algorithm performs best in various blind denoising algorithms.