Research On Image De-raining Based On Multi-Feature Fusion And Self-Calibrated Convolution

Rain often hinders outdoor image acquisition,leading to issues like color distortion and low definition caused by weather factors such as fog and thunderstorms.Meeting the increasing demand for high-quality images requires advancements in image rain removal algorithms.This dissertation focuses on the systematic exploration of image rain removal using multifeature fusion and self-correcting convolution,leveraging multi-scale features.To address the information loss problem in the pixel-by-pixel expansion filtering rain removal algorithm,we enhance the EncoderDecoder convolutional neural network.Our approach incorporates a dense residual module that gradually extracts multi-scale features from shallow to deep layers,emphasizing important image features with the SENet attention module.Additionally,we introduce the Multi-conv module between the encoder and the decoder,combining packet convolution,hole convolution,and Mask convolution to expand the model’s perception range and capture features across various scales.To further improve expressiveness,generalization,and reduce network errors,we propose an image raining algorithm based on self-correcting convolutional layers.This method introduces self-calibrating convolutional layer modules in the last layer of each encoder’s four convolutional modules,which is experimentally proven effective.For model training,we utilize four artificially synthesized datasets:Rain100L,Rain100 H,DIDMDN-Data,and DDN-Data.Real-world rainy datasets are used to evaluate the model’s generalization ability.Compared to recent deep learning-based de-raining algorithms,our proposed algorithm demonstrates superior performance,as indicated by improved PSNR and SSIM metrics,resulting in visually enhanced images.Furthermore,our de-raining algorithm showcases strong generalization capabilities by effectively addressing image dehazing problems.Finally,based on the rain removal algorithm proposed in this dissertation,we design an end-to-end B/S architecture image enhancement system.This system significantly improves the clarity and visibility of rainy or foggy images.Its user-friendly interface offers functionalities such as user interaction,image input,weather classification,image enhancement,image output,and result saving,enhancing user convenience.