As training billion-scale transformers becomes increasingly common, employing multiple distributed GPUs along with parallel training methods has become a standard practice. However, existing transformer designs suffer from significant communication overhead, especially in Tensor Parallelism (TP), where each block’s MHA–MLP connection requires an all-reduce communication. Through our investigation, we show that the MHA-MLP connections can be bypassed for efficiency, while the attention output of the first layer can serve as an alternative signal for the bypassed connection. Motivated by the observations, we propose FAL (First Attentions Last), an efficient transformer architecture that redirects the first MHA output to the MLP inputs of the following layers, eliminating the per-block MHA-MLP connections. This removes the all-reduce communication and enables parallel execution of MHA and MLP on a single GPU. We also introduce FAL+, which adds the normalized first attention output to the MHA outputs of the following layers to augment the MLP input for the model quality. Our evaluation shows that FAL reduces multi-GPU training time by up to 44%, improves single-GPU throughput by up to 1.18×, and achieves better perplexity compared to the baseline GPT. FAL+ achieves even lower perplexity without increasing the training time than the baseline. Codes are available at: https://casl-ku.github.io/FAL/

Deep learning has recently made an impressive advance in ship detection in Synthetic Aperture Radar (SAR) images. In spite of this advancement, conventional deep detection networks often suffer from speckle noise that inherently occurs in SAR images. However, despeckling researches have focused only on improving the visual quality of the SAR images. Despeckling without considering subsequent task may cause loss of semantic information and result in performance degradation. In this letter, we propose a deep cascade framework for noise-robust SAR ship detection that sequentially performs despeckle and detection. We effectively train our cascade network using pseudo SAR images with SAR-like structures and additional detection annotations. We also propose semantic conservative loss that allows these two tasks to cooperate with each other. Experimental results including comparisons to previous methods and extensive ablation studies show the effectiveness of our proposed method.

Motion blur, which disturbs human and machine perceptions of a scene, has been considered an unnecessary artifact that should be removed. However, the blur can be a useful clue to understanding the dynamic scene, since various sources of motion generate different types of artifacts. Motivated by the relationship between motion and blur, we propose a motion-aware feature learning framework for dynamic scene deblurring through multi-task learning. Our multi-task framework simultaneously estimates a deblurred image and a motion field from a blurred image. We design the encoder-decoder architectures for two tasks, and the encoder part is shared between them. Our motion estimation network could effectively distinguish between different types of blur, which facilitates image deblurring. Understanding implicit motion information through image deblurring could improve the performance of motion estimation. In addition to sharing the network between two tasks, we propose a reblurring loss function to optimize the overall parameters in our multi-task architecture. We provide an intensive analysis of complementary tasks to show the effectiveness of our multi-task framework. Furthermore, the experimental results demonstrate that the proposed method outperforms the state-of-the-art deblurring methods with respect to both qualitative and quantitative evaluations.

Convolutional neural networks (CNNs) have facilitated substantial progress on various problems in computer vision and image processing. However, applying them to image fusion has remained challenging due to the lack of the labelled data for supervised learning. This paper introduces a deep image fusion network (DIF-Net), an unsupervised deep learning framework for image fusion. The DIF-Net parameterizes the entire processes of image fusion, comprising of feature extraction, feature fusion, and image reconstruction, using a CNN. The purpose of DIF-Net is to generate an output image which has an identical contrast to high-dimensional input images. To realize this, we propose an unsupervised loss function using the structure tensor representation of the multi-channel image contrasts. Different from traditional fusion methods that involve time-consuming optimization or iterative procedures to obtain the results, our loss function is minimized by a stochastic deep learning solver with large-scale examples. Consequently, the proposed method can produce fused images that preserve source image details through a single forward network trained without reference ground-truth labels. The proposed method has broad applicability to various image fusion problems, including multi-spectral, multi-focus, and multi-exposure image fusions. Quantitative and qualitative evaluations show that the proposed technique outperforms existing state-of-the-art approaches for various applications.

Regularization-based image restoration is one of the most powerful tools in image processing and computer vision thanks to its flexibility for handling various inverse problems. However, designing an optimal regularization function still remains unsolved since natural images and related scene types have a complex structure. In this paper, we present a general and principled framework, called deeply aggregated alternating minimization (DeepAM). We design a convolutional neural network (CNN) to implicitly parameterize the regularizer of the alternating minimization (AM) algorithm. Contrary to the conventional AM algorithm based on a point-wise proximal mapping, the DeepAM projects intermediate estimate into a set of natural images via deep aggregation. Since the CNN is fully integrated into the AM procedure, all parameters can be jointly optimized through end-to-end training. These properties enable the DeepAM to converge with a small number of iterations, while maintaining an algorithmic simplicity. We show that the DeepAM outperforms state-of-the-art methods, including nonlocal-based methods, Plug-and-Play regularization, and recent data-driven approaches. The effectiveness of our framework is demonstrated in a variety of image restoration tasks: Guassian denoising, deraining, deblurring, super-resolution, color-guided depth upsampling, and RGB/NIR restoration.