Research On Image Specific Target Deblurring Algorithm

Image deblurring is an important research task in image restoration work.However,most of the current research on image deblurring focus on the whole image,but in real life,people often want to deblur specific objects in the image,such as human face,pedestrains,signage and license plate,etc.There is no need to deblur the background of the image.In this paper,we study the deblurring algorithm for specific objects mainly for pedestrains.The main work content and innovation are as follows:A generative adversarial network for pedestrian deblurring(Pedestrain Deblurring Network,PDNet)is proposed,which can better deblur the pedestrian in the image.The network consists of a generator and a discriminator.Among them,the generator adopts a feature pyramid backbone structure,which adopts the Inception-Res Net network framework,which is superior to other lightweight frameworks such as Moblile Net in performance;the discriminator adopts a dual-scale structure,and the global discriminator is based on the overall contour information in the image,and local discriminator focus on different detail information.This network is nearly twice as fast as other networks in terms of operation rate while having fewer parameters.Based on the above network structure,this paper proposes the following two algorithms to focus on deblurring pedestrians in images.(1)We design and implement a pedestrian target deblurring algorithm based on the target detection network(Target Detection Pedestrain Deblurring Network,TD-PDNet):The network detect pedestrians in the image,and then add the coordinates of the detection bounding box and the blurred image into the deblurring network,the weight is then added to the pedestrian part,which is similar to the attention mechanism,and the network’s attention is focused on the pedestrian part in the frame.Therefore,the task of deburring the pedestrain in the image is realized through the network training.The algorithm achieves a PSNR index of 29.63 and an SSIM index of 0.942 under the Go Pro dataset,and obtains a good visual deblurring effect.(2)A multi-scale convolution-based pedestrian deblurring end-to-end network algorithm(Multi-Scale Convolution Pedestrain Deblurring Network,MSConv-PDNet)is proposed: a multi-scale convolution feature fusion algorithm is added to the generator,and the module includes dilated convoltion,discarded convolution,and channeltransformed convolution.Among them,the dilated convolution is used to expand the local receptive field;the discarded convolution is used to regularly discard some data;the channel-transformed convolution is used to regularly disrupt the channel.And the pedestrian is weighted in loss backpropagation,which prompts the network to focus on extracting pedestrian features,so as to realize the key deblurring task of pedestrians.And we select all the images of the Go Pro test set,and mark the pedestrians in them to test the deblurring effect of this part.The PSNR index of the pedestrian part reaches40.45,the SSIM index reaches 0.992,and a good visual deblurring effect is also obtained.The PSNR index and SSIM on the HIDE dataset are 29.3 and 0.924 respectively,and a better deblurred visual effect map is obtained.In summary,the two PDNets proposed in this paper have achieved good results under both objective and subjective indicators,so as to verify the performance advantages of the algorithm,and a method is proposed to calculate the objective indicators of local image information.