Research On Lightweight Image Denoising Methods For Urban Security Surveillance Scenarios

The intelligent development trend of the modern urban security industry has given rise to a large number of intelligent security needs,however,the image video data quality in real security surveillance scenarios is difficult to guarantee,which seriously affects the execution effect of visual tasks in existing security surveillance scenarios,so how to solve the image noise problem in security surveillance scenarios has attracted extensive attention from many researchers.In recent years,image denoising methods based on deep learning have emerged,showing better denoising effects than traditional denoising methods to a certain extent,but the number of parameters and computational costs of existing deep denoising models is too large,and the complex and variable environments involved in real security surveillance scenes limit the denoising ability of deep models for real security surveillance images.Therefore,this paper proposes a lightweight image denoising method for real urban security surveillance scenes.Specifically,the research content of this paper is divided into the following two parts:(1)Due to a large number of model parameters and computational cost,the existing depth denoising methods are limited to be applied in security edge devices with limited computational performance.In this method,a uniform cascade structure is designed based on the Dense Net structure with residual connection to achieve efficient noise feature extraction.To achieve lightweight and efficient model parameters and computational cost,this paper investigates an image denoising lightweight method to adapt to multi-task requirements based on existing model lightweighting methods,and explores two lightweight implementation methods in this module.In this paper,two lightweight implementations are explored in this method,namely,a lightweight module based on depth-separable convolution and a lightweight module based on linear transformation operation,which can achieve different degrees of model parameter lightweight,and its plug-and-play design also allows researchers to choose different lightweight modules more flexibly to adapt to different denoising tasks.(2)To address the problem that existing depth denoising models are difficult to achieve efficient denoising in the face of noisy images of real security surveillance scenes,this paper organically incorporates a pixel shuffle sampling module into the denoising model mentioned in the previous study and explores a lightweight image denoising method based on the pixel shuffle sampling module.By studying the difference between spatial regional distribution of noise and simulated noise in real security scenes,the pixel blending algorithm for reducing the regional correlation of spatial distribution of real noise is studied,and the effective removal effect of the model on real security noise is realized;in addition,to address the problem that the existing image denoising dataset involves insufficient elements of real security scenes and it is difficult to train a deep denoising network for security surveillance scenes,the paper constructs the security surveillance scene image dataset City1000,which effectively helps the tasks of image denoising,image de-drainage,image de-fogging and image super-resolution reconstruction for security surveillance scenes.In this paper,we conducted extensive experiments on public datasets such as BSDS500 and Set12,verified that the denoising model proposed in this paper can achieve a more lightweight and effective image denoising goal by comparing with each existing denoising model,and explored the enhancement effect of each module on the overall model through ablation experiments.