Search

Deep convolutional neural networks (DCNNs) have exhibited excellent feature extraction and detail reconstruction capabilities for single image super-resolution (SISR). Nevertheless, most previous DCNN-based methods do not fully utilize the complementary strengths between feature maps, channels, and pixels. Therefore, it hinders the ability of DCNNs to represent abundant features. To tackle the aforementioned issues, we present a Cascaded Visual Attention Network for SISR called CVANet, which simulates the visual attention mechanism of the human eyes to focus on the reconstruction process of details. Specifically, we first designed a trainable feature attention module (FAM) for feature-level attention learning. Afterward, we introduce a channel attention module (CAM) to reinforce feature maps under channel-level attention learning. Meanwhile, we propose a pixel attention module (PAM) that adaptively selects representative features from the previous layers, which are utilized to generate a high-resolution image. Satisfactory, our CVANet can effectively improve the resolution of images by exploring the feature representation capabilities of different modules and the visual perception properties of the human eyes. Extensive experiments with different methods on four benchmarks demonstrate that our CVANet outperforms the state-of-the-art (SOTA) methods in subjective visual perception, PSNR, and SSIM.The code will be made available https://github.com/WilyZhao8/CVANet. [ABSTRACT FROM AUTHOR]

Capturing all-in-focus images with 3D scenes is typically a challenging task due to depth of field limitations, and various multi-focus image fusion methods have been employed to generate all-in-focus images. However, existing methods have difficulty simultaneously achieving real-time and superior fusion performance. In this paper, we propose a region- and pixel-based method that can recognize the focus and defocus regions or pixels by the neighborhood information in the source images. The proposed method can obtain satisfactory fusion results and achieve improved real-time performance. First, a convolutional neural network (CNN)-based classifier generates a coarse region-based trimap quickly, which contains focus, defocus and boundary regions. Then, precise fine-tuning is implemented at the boundary regions to address the boundary pixels that are difficult to discriminate by existing methods. Based on a public database, a high-quality dataset is constructed that provides abundant precise pixel-level labels, so that the proposed method can accurately classify the regions and pixels without artifacts. Furthermore, an image interpolation method called NEAREST_Gaussian is proposed to improve the recognition ability at the boundary. Experimental results show that the proposed method outperforms other state-of-the-art methods in visual perception and object metrics. Additionally, the proposed method has 80% improved to the conventional CNN-based methods. [ABSTRACT FROM AUTHOR]

Convolutional neural network (CNN)-based self-supervised visual representation learning (SSL) is a long-standing problem achieving significant successes with traditional handcrafted pretext tasks and contrastive learning. However, existing SSL methods typically suffer from high computational overhead and poor performance due to sluggish convergence speeds and poor detail extraction capabilities. In this work, to address these issues and improve the robustness, we provide a new self-supervised architecture for incorporating a single-branch backbone with hybrid tasks into the representation learning process. Specifically, our method takes advantage of features from both intra- and inter-images by using discrete montage images. Then a single backbone with a novel A daptive D ecouple C onfusion (ADC) module is proposed to improve the feature extraction capabilities and alleviate the confusion regions in montage images. Besides, both concatenated discrete vectors and patch-based global average pooled vectors in latent space are utilized to learn local detailed features and maintain semantic consistency simultaneously. Moreover, our method is optimized by hybrid tasks and enjoys faster convergence speed due to these improvements. Extensive experiments on several datasets demonstrate the effectiveness and robustness of our method. The proposed method has 2.0% improved in linear classification to the conventional single-branch methods. [Display omitted] • A new framework SSH that leverages concatenated discrete and PGAP vectors is proposed. • An A daptive D ecouple C onfusion (ADC) module is proposed to avoid degenerate performance in existing jigsaw-based methods. • We demonstrate the advantages of the above contributions through extensive experiments. [ABSTRACT FROM AUTHOR]

Capturing the comprehensive information of various sizes and shapes of images in the same convolution layer is typically a challenging task in computer vision. There are two main kinds of methods for capturing those features. The first uses the inception structure and its variants. The second utilizes larger convolution kernels on specific layers or stacks with more convolution blocks. However, these methods can result in computationally intensive or vanishing gradients. In this paper, to accommodate feature distributions with different sizes, shapes and reduce computational cost, we propose a width- and depth-aware module named the WD-module to match feature distributions. Moreover, the proposed WD-module consumes less computational cost and parameters compared with traditional residual convolution layers. To verify the effectiveness of our proposed method, a size- and shape-aware backbone network named S2A-Net was built, which was obtained by stacking the WD-modules. By visualizing heat maps and features, the proposed S2A-Net can adapt to objects with different sizes and shapes in visual recognition tasks and learn more comprehensive characteristics. Experimental results show that the proposed method has higher accuracy in image recognition and outperforms other state-of-the-art networks with the same numbers of layers. • A width-aware and deep-aware module is build. • A size-aware and shape-aware network for visual recognition is build. • Compared with other state-of-the-art methods, the proposed method can achieve better results while consume less parameters and FLOPs. [ABSTRACT FROM AUTHOR]