An Image Deraining Algorithm Driven By Matching Task

In severe weather conditions,such as rainstorm,haze or snow,images captured by outdoor camera often are noisy,blurred and lack of feature information.These problems will lead to the performance degradation of many computer vision tasks,which rely on high-quality inputs,especially image matching which based on local feature.As a basic visual task,image matching plays an important role in visual applications.Nowadays,the mainstream deraining algorithms are usually generating a clean and clear image without considering the preservation of local features.Consequently,it is important to design a algorithm of removing rain,snow or haze which can preserve more feature information.Besides,rainy days are more common in daily life.Therefore,we focus on an image deraining algorithm which can retain more feature information.In order to retain more feature information,our proposed algorithm based on scale space,directly generate derained Gaussian images at different scales.After rain removal,scale space constructed by Gaussian images can be applied to image matching task to improve the accuracy of matching.In this paper,we design a residual deraining algorithm based on encoder-decoder architecture.In details,an encoder-decoder based-on module is applied to generate the maximum scale derained Gaussian image,and several residual network based-on modules are used to generate derained Gaussian images at the other scales.Different modules are connected in cascades.In particular,we use a step-by-step training method: when the training of rain-removal Gaussian images at one scale is completed,its model parameters will be frozen.And the deraining module at the next scale is trained on the basis of these parameters.In addition,in order to remove additional gradient information caused by rain and preserve more accurate feature information,each model will be fine-tuned at the end of the training using the Structural Similarity Index Measurement(SSIM)loss of the gradient of Gaussian images.In this paper,the derained Gaussian images are used in Scale Invariant Feature Transform(SIFT)to evaluate performance.We compare our algorithm with other deraining algorithms in both quantitative and qualitative tests on five open datasets.Results show that our algorithm can retain more feature information,which improve the robustness and accuracy of image matching under rainy conditions.