Research On Multi-scale Feature Equalization Architecture Of Deep Neural Networks And Applications

Multi-scale feature fusion plays an important role in many classical deep learning network models.Multi-scale means observing signal data at different granularities.When the granularity is larger and sparser,the network can learn the overall trend and obtain global high-level features;When the granularity is smaller and denser,the network can learn more details and obtain local underlying characteristics.The features learned at different granularity are integrated through a unique spatial channel equalization method,which makes the network pay attention to texture and structure at the same time,thus ensuring the high efficiency and robustness of the model.In the research,we will mainly focus on the multi-scale feature fusion technology in the deep neural network,combined with the image inpainting problem in the field of computer vision and the problem of machine intelligence diagnosis in the era of Industry 4.0,aiming to innovate the network structure and algorithm.At the same time,we also give a detailed overview of the connotation and development of multi-scale feature fusion technology.The classical network architectures that are currently popular and have reference significance for this study are introduced.Specifically,the research content of this paper can be summarized as follows:(1)For the problem of data augmentation for datasets with partially damaged or missing images,we first classify them into the field of image inpainting.The difficulty of the problem is how to ensure the effect of image inpainting and reuse the decoder part of the image inpainting model at the same time for data augmentation.Therefore,we shifted our attention from the confrontation generation network to the auto-encoder and proposed a multi-scale feature equalization fusion multi-level variational auto-encoder network(Multi-level VAE For Image Inpainting,MVAE-FI),which changes the single-level structure of the traditional autoencoder and improves the generation ability of the model.A multi-scale feature balance fusion module is introduced between the encoder and decoder,which makes the model pay more attention to image detail texture and overall semantic features.The experimental results show that MVAE-FI has satisfactory results both in the visual effect of image inpainting and quantifiable evaluation indicators.At the same time,the goal of data augmentation can be achieved by multiplexing the decoder in MVAE-FI.(2)The diagnostic capability of a model in the field of machine intelligence diagnosis is positively correlated with its scale.However,a model with a large amount of computation is difficult to practically applied to an industrial production environment.To solve the problem of how to balance the diagnostic capability of a model and its scale,we first propose a multi-scale feature fusion densely connected deep network(MSDC-NET),which uses the multi-scale feature fusion technology to expand the receptive field of the model,and avoids the gradient explosion of the model through the dense connection technology.Finally,the overall network architecture and local residual components of the MSDC-NET model are well-designed,and the performance of the three types of data sets far exceeds the current mainstream machine intelligence diagnosis algorithms.In addition,we perform knowledge distillation on the lightweight CNN model based on the MSDC-NET model with a high-precision classification level,which greatly improves the diagnostic accuracy.Since the computational and storage device requirements required by the lightweight CNN model are much lower than those of the deep network,it plays a positive role in the promotion of machine intelligence diagnosis in industrial practice.