Research On A Single Image Rain Removal Method Based On Deep Learning

Images taken in rainy weather will be affected by bad weather,resulting in quality degradation,and the background information of the image will be blocked by rain streaks,and the imaging visibility will be significantly reduced.In normal rainy weather,advanced computer vision tasks such as object detection,semantic segmentation and so on,which is familarly not performed correctly because they all require high-quality input images.Therefore,it is highly important to propose a single image deraining algorithm with a lightweight network model,fast execution speed,and high output derained image quality,to restore an image with rain streaks to an image without rain streaks.Although the previous methods of removing rain streaks from a single image have a good effect,they are frequently only suitable for removing rain streaks in certain fixed scenes.First of all,most of the traditional model-based single image rain removal methods are based on statistical analysis of rain images to obtain some assumptions that could separate rain streaks.However,assumptions only be used correctly in specific rainy scenes situations,and not in all rainy images.Therefore,when it is directly used in the real rain images to remove the rain,the result of the derain images can frequently not accurately separate the rain streaks layer and the background no rain layer.Secondly,the currently proposed single image rain removal algorithm based on deep neural network uses the artificially added noise synthetic rain image and the image pair composed of the original rain-free image for training when the rain streaks removal network is trained.Which could avoid the use of physical model of rain and do not require adding priors.However,when these rain removal methods remove the rain streaks features of the rain image,they will remove the background texture details of the rain image,resulting in the background of the obtained rain image being too smooth and blurred.Based on the shortcomings of the above rain removal methods,this paper proposes two different single image rain removal methods.The specific overview is as follows:Firstly,according to the time sequence proposed by the rain removal methods,a variety of proposed single image rain removal methods are summarized.The advantages and disadvantages of the existing rain removal methods are expounded.Then,in view of the problem that the existing deep image deraining algorithm has plenty of network layers,high complexity,and difficulty in training,this paper proposes a new Detail-guided Efficient Channel Attention referred to as DECA.Which could adaptively extract global information and detail information in rain images.Moreover,the proposed DECA module is extremely lightweight and could support densely embedded deraining networks.The experiments results show that adding DECA to the rain removal network could better protect the global and local features in the original rain image,making the background of the rain removal image look better,and the convergence speed of the rain removal network is faster during training.Based on the Detail-Guided Attention Module(DECA),a Residual Rain Streaks Removal Network(RSRN)is proposed.RSRN is able to effectively separate the global information and detail features in the rain image when recognizing rain streaks,thereby extracting accurate rain streaks.In addition,the Background Detail Recovery Network under the global view is also proposed,which constrains the background detail reconstruction under the global information and avoids the rain streaks information from being incorrectly reconstructed.Finally,a lightweight multi-stage image deraining algorithm is proposed for the problems of too large network model,too many parameters,and low operating efficiency,named Light Multi-stage Deraining Network,or LMDN for short.The algorithm is based on the newly proposed lightweight network Ghost design idea,and all ordinary convolutions are implemented using Ghost modules instead.Then use the multi-stage idea to gradually and gradually remove the rain streaks to restore the rainfree background image.The first two stages are based on a lightweight encodingdecoding network structure,and the third stage uses the original resolution residual network and multi-scale context aggregation repair module.The composition is designed to compensate for the background image details lost during the first two stages of de-raining.An effective supervised attention mechanism is used to connect the three stages,and its function is to pass on the features of the rainy image extracted in the current stage that can best help complete the image rain removal,as the input of the next stage rain removal network,so as to improve the rain removal accuracy and efficiency.