Lightweight pixel volume deblurring network: enhanced video deblurring via efficient architecture optimization

A fast deblurring network, based on a high-performance convolutional network and pixel volume, is proposed to address the limitations of existing video deblurring algorithms, which often overly emphasize inter-frame information, leading to high algorithmic complexity. First, high-performance convolutional networks are utilized to prune the deblurring network, thereby reducing both the number of model parameters and computational complexity. To address the increased network computational complexity resulting from the extensive use of traditional two-dimensional convolutional layers, depthwise over-parameterized convolutions are employed to replace traditional convolutions. This substitution significantly reduces computational complexity without compromising the network’s structure and performance. In addition, the Charbonnier loss function is used to approximate the mean absolute error (MAE) loss function to alleviate the over-smoothing problem. At the same time, the problem of non-differentiability of the MAE loss function at zero is solved by adding a constant, to enhance the visual quality of video images. Experimental results demonstrate that the proposed method delivers superior deblurring performance. Compared with the baseline pixel volume deblurring network framework, our method achieves a significant reduction in model complexity, demonstrating 28.73% fewer parameters and 59.96% lower floating-point operations, underscoring its theoretical significance.