A Plug-and-Play Deraining Network For Object Detection

Since existing deep object detection methods are trained on clear images,their performances degrade under poor imaging environment eg rainy conditions due to low contrast,color distortion,blurriness and imbalanced illumination.Thus,in-depth study of detection under poor imaging environment is urgently required.At present,deep learning have achieved more excellent performance than traditional methods on different images processing tasks,such as rain removal,low light enhancement and defogging.However,most CNN-based deraining models have large number of parameters which make them difficult to be embedded into detection networks which restrict the application of target detection algorithm in harsh imaging environment.Moreover,these methods only focus on the improvement of subjective and objective performance indicators of the image,and do not consider the impact with the following object detection network.This leads to the detection performance is still much lower than the one achieved with clear inputs.To address the above issues,this work present a plug-and-play network to improve the performance of detection network under rainy weather based on convolutional neural network.The main content and innovations are follows:(1)The work proposes a light weight and end-to-end network by using dilated convolution and recursive structure for image deraining.Our network can extract spatial information at different scales while has less than 40K parameters.(2)To further improve the performance of subsequent detection network,a novel loss is designed to train the network.Specifically,this work introduce the SSIM as a part of loss function to force the network to generate sharp results.The L1 perceptual loss is also adopt to link deraining and detection through existing high-level feature extraction.So as the detection module participates in the evaluation of the deraining image.When jointly applying the two losses,the output image of the rain removal network can effectively solve the degradation problem of the rainy images and make its target detection performance greatly improved.Experiments in both synthetic and real-world rainy images demonstrate that proposed algorithm can achieve the removal of rain streaks,image detail reveal and sharpness.It can significantly improve the rainy images quality with better computational efficiency and less parameters.Moreover,it can greatly improve the performance of object detection under rainy weather imaging environment.