Improved Deep Multi-Patch Hierarchical Network For Handling Saturation In Image Deblurring

Abstract

The most active study topic in the field of computer vision is dealing with saturation in image deblurring. The Deep Multi Patch Framework (DMPHN) is primarily utilized in dynamic scene deblurring, and its fine-to-coarse hierarchical representation results in lower processing costs. The contribution of the framework in gathering and processing local residual information of blur at the courser levels in the hierarchy makes the dynamic scene deblurring successful. The existing framework uses the traditional CNN with the convolutional filter, which is a generalized linear model (GLM) that goes along with a nonlinear activation function to scan the input. When the samples of latent concepts are not linearly separable then the linear CNN cannot abstract the good representations. Because saturation in an image contradicts the linear properties of the blur model, the present Deep Multi Patch architecture with linear CNN is unable to collect the necessary local information of blur in an image. In CNN, the higher-level layer maps to the larger regions of the input image by combining the lower-level concepts from each local patch. Due to the framework’s incapacity to retain information at lower levels, noticeable ringing artifacts and poor deblurring performance resulted. We designed a nonlinear function approximator for this section, which was influenced by the good abstraction capabilities of ”Network-in-Network”. The nonlinear features of saturated pixels and blur are successfully captured by this upgraded DMPHN with a nonlinear function approximator. Compared to the current DMPHN model, the suggested approach greatly reduced the number of trainable parameters while improving the time cost. On the GoPro, Synthetic, and Natural Saturated Blurred Image datasets, the upgraded DMPHN network achieves superior deblurring performance with better PSNR and SSIM than the state-of-the-art.