Polarization Information Restoration of Underwater Images Based on Deep Neural Network
Objective Underwater imaging is an important method for exploring oceans,lakes,and other underwater environments,and it is of significance for many fields such as coastal defense,ocean exploration,underwater rescue,and aquaculture.However,there are many suspended particles in the actual water environment,which will scatter and absorb the signal light of the target.Therefore,images obtained by underwater imaging often feature image quality degradation,such as serious contrast reduction and serious detail loss.Due to the differences in polarization characteristics between the signal light and the backscattered light,polarization imaging technology is introduced to underwater imaging and polarization information is employed to suppress scattered light and enhance signal light,which can make up for the shortcomings of detection effects restricted by the environment.Although existing polarization-based descattering methods for underwater imaging can enhance image contrast and improve image quality,these methods only focus on intensity information and ignore polarization information restoration,resulting in a loss of polarization information.In fact,the degree of linear polarization(DoLP)and polarization angle(AoP)among other polarization information can reflect the polarization characteristics of the target.Meanwhile,they are adopted to distinguish targets based on different polarization states and have important applications in underwater detection and recognition.Therefore,we propose a neural network based on the channel attention mechanism to extend the function of underwater polarization imaging by restoring polarization information.Methods The proposed method mainly utilizes a convolutional neural network to restore the polarization information.The network is mainly composed of three parts,including shallow feature extraction module(SFE),a series of residual dense modules(RDBs),and channel attention-based global feature fusion module(CAGFF).Specifically,SFE employs U-Net and two convolutional layers as feature extraction module to extract shallow features containing polarization features of input images.Subsequently,the shallow features with rich polarization information extracted by SFE are input into a series of RDBs,which are mainly composed of dense connections and residual connections of convolutional layers.A series of RDBs outputs are then fed into the CAGFF which consists of a channel attention module and a convolutional layer.Finally,the polarization informed content and style loss(CSL)is designed to train the network,which employs intensity images to calculate the content loss and adopts polarization information to calculate style loss.Results and Discussions The results of polarization imaging experiments of different objects in underwater environments show that our method can restore polarization information and improve underwater imaging quality.In ablation experiments,the four contributions of the network are removed step by step to verify the effectiveness of each contribution.By visual comparison,the network with all improvements accurately recovers both intensity image and polarization information(Fig.5).Removing any component will degrade the restored results and blur details.Quantitatively,our network yields the highest peak signal to noise ratio(PSNR)value among different network structures(Table 2).Compared with other representative methods of underwater image enhancement,the proposed method has the best performance in intensity,DoLP,and AoP images(Fig.6).The PSNRs of two different scenes illustrate that the proposed method has advantages in the reconstruction of DoLP and AoP images.Additionally,several groups of experiments on different parameters are conducted to verify the influence of these parameters.First.the number of channel attention modules is determined by comparing the network performance with different values(Fig.7).Then by the comparative experiment of RDB parameters,we find that the values of RDB parameters are positively correlated with the network performance(Fig.8).Finally,the weight of the loss function is fully discussed(Fig.9).Conclusions Aiming to remove the scattered light and restore the polarization information,we propose a neural network based on the channel attention mechanism.Based on the residual dense network,our network mainly has four contributions,which are polarization input,SFE with U-Net,CAGFF,and polarization informed CSL function.By performing the above contributions,the network can efficiently extract and utilize the polarization features of different levels to restore the polarization information.Meanwhile,we build a dataset for underwater polarimetric images,in which the input images and ground truth images are respectively obtained in turbid water and clear water.Based on this dataset,we carry out a series of experiments on different objects in underwater environments.Ablation experimental results show that our contributions to the network are effective,and removing anyone will degrade the results.Compared with other underwater polarization imaging methods,the results show that our method can suppress the influence of scattered light on polarization imaging and significantly improve the image contrast and clarity.Specifically,our method can successfully restore the DoLP and AoP to expand the function of underwater polarization imaging.
polarimetric imagingunderwater imagingpolarization information restorationdeep learningattention mechanism
NETL
NSTL
万方数据
