基于渐进式特征增强网络的超分辨率重建算法

为了在图像重建质量和网络参数之间取得较好的平衡,本文提出一种基于渐进式特征增强网络的超分辨率（Super-Resolution,SR）重建算法。该方法主要包含两个模块:浅层信息增强模块和深层信息增强模块。在浅层信息增强模块中,首先利用单层卷积层提取低分辨率（Low-Resolution,LR）图像的浅层信息,再通过我们设计的多尺度注意力块来实现特征的提取和增强。深层信息增强模块先利用残差学习块学习图像的深度信息,然后将得到的深层信息通过设计的多尺度注意力块来获得增强后的深层多尺度信息。最后我们利用跳转连接的方式将首层得到的浅层信息和深层多尺度信息进行像素级相加得到融合特征图,再对其进行上采样操作,得到最终的高分辨率（High-Resolution, HR）图像。实验结果表明,相比于一些主流的深度学习超分辨率方法,本文方法重建得到的图像无论是主观效果还是客观指标,都取得了更好的效果。      

Multi-scale convolutional attention network for lightweight image super-resolution

Convolutional neural network (CNN) based methods have recently achieved extraordinary performance in single image super-resolution (SISR) tasks. However, most existing CNN-based approaches increase the model’s depth by stacking massive kernel convolutions, bringing expensive computational costs and limiting their application in mobile devices with limited resources. Furthermore, large kernel convolutions are rarely used in lightweight super-resolution designs. To alleviate the above problems, we propose a multi-scale convolutional attention network (MCAN), a lightweight and efficient network for SISR. Specifically, a multi-scale convolutional attention (MCA) is designed to aggregate the spatial information of different large receptive fields. Since the contextual information of the image has a strong local correlation, we design a local feature enhancement unit (LFEU) to further enhance the local feature extraction. Extensive experimental results illustrate that our proposed MCAN can achieve better performance with lower model complexity compared with other state-of-the-art lightweight methods.     

Recursive residual atrous spatial pyramid pooling network for single image deraining

In recent years, removing rain streaks from a single image has been a significant issue for outdoor vision tasks. In this paper, we propose a novel recursive residual atrous spatial pyramid pooling network to directly recover the clear image from rain image. Specifically, we adopt residual atrous spatial pyramid pooling (ResASPP) module which is constructed by alternately cascading a ResASPP block with a residual block to exploit multi-scale rain information. Besides, taking the dependencies of deep features across stages into consideration, a recurrent layer is introduced into ResASPP to model multi-stage processing procedure from coarse to fine. For each stage in our recursive network we concatenate the stage-wise output with the original rainy image and then feed them into the next stage. Furthermore, the negative SSIM loss and perceptual loss are employed to train the proposed network. Extensive experiments on both synthetic and real-world rainy datasets demonstrate that the proposed method outperforms the state-of-the-art deraining methods.     

基于改进U-Net网络的腺体细胞图像分割算法

针对腺体图像在自动分割过程中由于多尺度目标和信息丢失影响导致准确率降低的问题,文中采用了一种引入注意力模块的全卷积神经网络模型。该模型遵循编码器-解码器结构,在编码网络中用空洞残差卷积层代替原有的普通卷积层,并添加空洞金字塔池;再在解码网络中加入注意力模块,使模型输出高分辨率特征图,提高对多尺度目标的分割精度。实验结果表明,提出的网络模型参数少分割精度高,对腺体图像的平均分割精度高达89.7%,具有较好的鲁棒性。     

结合空洞可分离卷积的多尺度医学影像分割算法

传统U-Net语义分割模型在医学影像领域具有广泛的应用,但该模型的准确率受限于单一尺度的预测模式以及上下采样引起的信息丢失。针对上述问题,本文基于U-Net编码—解码架构以及空洞可分离卷积提出了一种高低层级信息丰富的多尺度医学影像语义分割算法,该算法由特征提取网络以及多尺度语义分割预测网络两部分构成。特征提取网络使用空洞可分离卷积和类残差块分别替换原U-Net中上、下采样以及卷积块,在增加感受野的同时使信息得到最大化的保留;提出一个通道注意力机制,强化目标核心特征的表达以及无关背景区域的抑制;在多尺度上挖掘带有图像级全局上下文的卷积特征,进一步提高分割性能。本文在采集的胚胎以及DRIVE数据集上进行仿真实验,其结果表明,与U-Net及其衍生模型相比该方法具有更高的准确率和鲁棒性。     

基于先验信息与密集连接网络的人脸超分辨率重建方法

图像超分辨率在医疗和安防等领域应用广泛,本文针对传统超分辨率重建(super-resolution reconstruction, SR)方法无法重建出边缘特征图像的不足,提出了一种先验信息与密集连接网络模型的重建方案,利用考虑输入统计信息的残差特征的不同组合,引入了多注意力模块,通过与主干网络结构协作,在不增加额外模块的情况下提高了网络性能。新提出的模型与现有复杂结构的技术(state-of-the-art, SOTA)模型相比,具有更好的性能。为了避免输入的身份特征会急剧漂移的问题,提出了一种基于先验信息引入注意力机制网络模块来分辨真实低分辨率(low resolution, LR)对应物的模型,这种模型在捕获运动噪声等方面具有优势。经实验验证得出,本文方法相比其他主流方法,在评价指标和主观可视化分析方面更具优势。     

基于轻量级金字塔密集残差网络的红外图像超分辨增强

现有的红外制导武器严重依赖操作手对目标的捕获,其捕获的精度与目标的纹理细节正相关.为了提升弱小区域的显示质量,满足现有导引头小型化、模块化、低成本的设计要求,本文设计了一种基于轻量级金字塔密集残差网络的图像增强模型,该模型在密集残差网络基础上通过密集连接层和残差网络来学习不同尺度图像之间的非线性映射,充分利用多尺度特征...     

基于多尺度特征残差学习卷积神经网络的视频超分辨率方法

本文提出了一种基于多尺度特征残差学习卷积神经网络的视频超分辨率方法,考虑到视频帧间的时空相关性,所提的方法采用由双三次插值预处理后的连续五帧视频作为卷积神经网络的输入,经由网络重建中间帧作为输出,依次按顺序重建直至获得整个高分辨率视频。本文所提出的卷积神经网络主要由多尺度特征提取、残差学习、亚像素卷积层、残差连接(skip-connection)四大部分组成,通过对视频的多尺度特征的提取获得更丰富的不同尺度特征和残差学习达到较好地恢复高频信息的目的。本文采用峰值信噪比（PSNR）和结构相似性指数（SSIM）作为损失函数优化网络。实验结果表明,本方法在平均评价指标上较其他方法均有一定的提升（PSNR +3.151dB,SSIM +0.102）,从主观评价上看可以有效地减少视频边缘模糊的现象。      

基于多尺度特征融合和注意力机制的医学图像分割网络

针对传统编解码结构的医学图像分割网络存在特征信息利用率低、泛化能力不足等问题,该文提出了一种结合编解码模式的多尺度语义感知注意力网络(multi-scale semantic perceptual attention network,MSPA-Net) 。首先,该网络在解码路径加入双路径多信息域注意力模块(dual-channel multi-information domain attention module,DMDA) ,提高特征信息的提取能力;其次,网络在级联处加入空洞卷积模块(dense atrous convolution module,DAC) ,扩大卷积感受野;最后,借鉴特征融合思想,设计了可调节多尺度特征融合模块 (adjustable multi-scale feature fusion,AMFF) 和双路自学习循环连接模块(dual self-learning recycle connection module,DCM) ,提升网络的泛化性和鲁棒性。为验证网络的有效性,在CVC-ClinicDB、ETIS-LaribPolypDB、COVID-19 CHEST X-RAY、Kaggle_3m、ISIC2017和Fluorescent Neuronal Cells等数据 集上进行验证,实验结果表明,相似系数分别达到了94.96%、92.40%、99.02%、90.55%、92.32%和75.32%。因此,新的分割网络展现了良好的泛化能力,总体性能优于现有网络,能够较好实现通用医学图像的有效分割。     

Gesture image recognition method based on DC-Res2Net and a feature fusion attention module

To extract decisive features from gesture images and solve the problem of information redundancy in the existing gesture recognition methods, we propose a new multi-scale feature extraction module named densely connected Res2Net (DC-Res2Net) and design a feature fusion attention module (FFA). Firstly, based on the new dimension residual network (Res2Net), the DC-Res2Net uses channel grouping to extract fine-grained multi-scale features, and dense connection has been adopted to extract stronger features of different scales. Then, we apply a selective kernel network (SK-Net) to enhance the representation of effective features. Afterwards, the FFA has been designed to remove redundant information in features by fusing low-level location features with high-level semantic features. Finally, experiments have been conducted to validate our method on the OUHANDS, ASL, and NUS-II datasets. The results demonstrate the superiority of DC-Res2Net and FFA, which can extract more decisive features and remove redundant information while ensuring high recognition accuracy and low computational complexity.     

Image super-resolution based on deep neural network of multiple attention mechanism

At present, the main super-resolution (SR) method based on convolutional neural network (CNN) is to increase the layer number of the network by skip connection so as to improve the nonlinear expression ability of the model. However, the network also becomes difficult to be trained and converge. In order to train a smaller but better performance SR model, this paper constructs a novel image SR network of multiple attention mechanism(MAMSR), which includes channel attention mechanism and spatial attention mechanism. By learning the relationship between the channels of the feature map and the relationship between the pixels in each position of the feature map, the network can enhance the ability of feature expression and make the reconstructed image more close to the real image. Experiments on public datasets show that our network surpasses some current state-of-the-art algorithms in PSNR, SSIM, and visual effects.     

Single image deraining using multi-scales context information and attention network

The existing deraining methods based on convolutional neural networks (CNNs) have made great success, but some remaining rain streaks can degrade images drastically. In this work, we proposed an end-to-end multi-scale context information and attention network, called MSCIANet. The proposed network consists of multi-scale feature extraction (MSFE) and multi-receptive fields feature extraction (MRFFE). Firstly, the MSFE can pick up features of rain streaks in different scales and propagate deep features of the two layers across stages by skip connections. Secondly, the MRFFE can refine details of the background by attention mechanism and the depthwise separable convolution of different receptive fields with different scales. Finally, the fusion of these outputs of two subnetworks can reconstruct the clean background image. Extensive experimental results have shown that the proposed network achieves a good effect on the deraining task on synthetic and real-world datasets. The demo can be available at https://github.com/CoderLi365/MSCIANet.     

基于语义分割的多目标违禁品识别算法

基于深度学习的语义分割算法可以实现安检违禁品自动识别,并获得违禁品的位置、类别及形状信息。但传统的语义分割算法在面对违禁品尺寸不一且目标多样的识别任务时表现较差。针对该问题,本文提出了一种基于语义分割技术的多目标违禁品识别算法。编码阶段,设计使用空洞空间金字塔卷积模块（Atrous Spatial Pyramid Convolution Block, ASPC）,提升网络对于特征图多尺度信息的挖掘能力。同时引入注意力机制,对ASPC模块的特征提取过程进行监督,进一步提升模块的特征提取能力。解码阶段,受U-Net模型启发,采用逐级上采样操作,同时加入1×1卷积实现通道降维,减少计算量,提升模型运行速度。实验结果显示,本文提出的算法在多目标违禁品识别任务中表现良好,平均交并比（mIoU）得分78.62,处理单张图片用时（Time）68ms。      

Underwater image super-resolution using multi-stage information distillation networks

Recently, single image super-resolution (SISR) has been widely applied in the fields of underwater robot vision and obtained remarkable performance. However, most current methods generally suffered from the problem of a heavy burden on computational resources with large model sizes, which limited their real-world underwater robotic applications. In this paper, we introduce and tackle the super resolution (SR) problem for underwater robot vision and provide an efficient solution for near real-time applications. We present a novel lightweight multi-stage information distillation network, named MSIDN, for better balancing performance against applicability, which aggregates the local distilled features from different stages for more powerful feature representation. Moreover, a novel recursive residual feature distillation (RRFD) module is constructed to progressively extract useful features with a modest number of parameters in each stage. We also propose a channel interaction & distillation (CI&D) module that employs channel split operation on the preceding features to produce two-part features and utilizes the inter channel-wise interaction information between them to generate the distilled features, which can effectively extract the useful information of current stage without extra parameters. Besides, we present USR-2K dataset, a collection of over 1.6K samples for large-scale underwater image SR training, and a testset with an additional 400 samples for benchmark evaluation. Extensive experiments on several standard benchmark datasets show that the proposed MSIDN can provide state-of-the-art or even better performance in both quantitative and qualitative measurements.     

Reference guided image super-resolution via efficient dense warping and adaptive fusion

Due to the limited improvement of single-image based super-resolution (SR) methods in recent years, the reference based image SR (RefSR) methods, which super-resolve the low-resolution (LR) input with the guidance of similar high-resolution (HR) reference images are emerging. There are two main challenges in RefSR, i.e. reference image warping and exploring the guidance information from the warped references. For reference warping, we propose an efficient dense warping method to deal with large displacements, which is much faster than traditional patch (or texture) matching strategy. For the SR process, since different reference images complement each other, and have different similarities with the LR image, we further propose a similarity based feature fusion strategy to take advantage of the most similar reference regions. The SR process is realized by an encoder–decoder network and trained with pixel-level reconstruction loss, degradation loss and feature-level perceptual loss. Extensive experiments on three benchmark datasets demonstrate that the proposed method outperforms state-of-the-art SR methods in both subjective and objective measurements.     

A new parallel hardware architecture for high-performance stereo matching calculation

In this paper, we propose a hardware (H/W) architecture to find disparities for stereo matching in real time. After analyzing the arithmetic characteristic of stereo matching, we propose a new calculating method that reuses the intermediate results to minimize the calculation load and memory access. From this, we propose a stereo matching calculation cell and a new H/W architecture. Finally, we propose a new stereo matching processor. The implemented H/W can operate at the clock frequency of 250 MHz at least in the FPGA (field programmable gate array) environment and produce about 120 disparity images per second for HD stereo images.     

基于门控卷积神经网络的图像超分辨重建算法北大核心CSCD

近年来,卷积神经网络被广泛应用于图像超分辨率领域。针对基于卷积神经网络的超分辨率算法存在图像特征提取不充分,参数量大和训练难度大等问题,本文提出了一种基于门控卷积神经网络(gated convolutional neural network, GCNN)的轻量级图像超分辨率重建算法。首先,通过卷积操作对原始低分辨率图像进行浅层特征提取。之后,通过门控残差块(gated residual block, GRB)和长短残差连接充分提取图像特征,其高效的结构也能加速网络训练过程。GRB中的门控单元(gated unit, GU)使用区域自注意力机制提取输入特征图中的每个特征点权值,紧接着将门控权值与输入特征逐元素相乘作为GU输出。最后,使用亚像素卷积和卷积模块重建出高分辨率图像。在Set14、BSD100、Urban100和Manga109数据集上进行实验,并和经典方法进行对比,本文算法有更高的峰值信噪比(peak signal-to-noise ratio,PSNR)和结构相似性(structural similarity,SSIM),重建出的图像有更清晰的轮廓边缘和细节信息。     

一种有效融合多尺度特征的图像语义分割方法

卷积神经网络在高级计算机视觉任务中展现出强 大的特征学习能力,已经在图像语义 分割任务 中取得了显著的效果。然而,如何有效地利用多尺度的特征信息一直是个难点。本文提出一 种有效 融合多尺度特征的图像语义分割方法。该方法包含4个基础模块,分别为特征融合模块(feature fusion module,FFM)、空 间信息 模块(spatial information module,SIM)、全局池化模块(global pooling module,GPM)和边界细化模块(boundary refinement module,BRM)。FFM采用了注意力机制和残差结构,以提高 融合多 尺度特征的效率,SIM由卷积和平均池化组成,为模型提供额外的空间细节信息以 辅助定 位对象的边缘信息,GPM提取图像的全局信息,能够显著提高模型的性能,BRM以残差结构为核心,对特征图进行边界细化。本文在全卷积神经网络中添加4个基础模块, 从而有 效地利用多尺度的特征信息。在PASCAL VOC 2012数据集上的实验结 果表明该方法相比全卷积神 经网络的平均交并比提高了8.7%,在同一框架下与其他方法的对比结 果也验证了其性能的有效性。     

利用多尺度卷积神经网络的图像超分辨率算法

单幅图像超分辨率问题是典型的图像反问题。近年来深度学习广泛应用于图像超分辨率重建。为提高超分辨率算法的性能,本文利用多尺度和残差训练的思想,提出一种利用多尺度卷积神经网络的图像超分辨率算法。该算法采用多尺度的卷积核及收缩--扩展的网络结构来提取图像多尺度的信息,并在网络结构中使用跳跃连接,以便更好的传递信息并弥补由于使用下采样和上采样而造成的图像细节信息的损失,来提高图像的重建质量。通过与其它算法的对比实验表明了本文算法不仅可以取得更好的性能,并且训练的收敛速度较快。      

Novel shrinking residual convolutional neural network for efficient accurate stereo matching

For stereo matching based on patch comparing using convolutional neural networks (CNNs), the matching cost estimation is highly dependent on the network structure, and the patch comparing is time consuming for traditional CNNs. Accordingly, we propose a stereo matching method based on a novel shrinking residual CNN, which consists of convolutional layers and skip-connection layers, and the size of the fully connected layers decreases progressively. Firstly, a layer-by-layer shrinking size model is adopted for the full-connection layers to greatly increase the running speed. Secondly, the convolutional layer and the residual structure are fused to improve patch comparing. Finally, the Loss function is re-designed to give higher weights to hard-classified examples compared with the standard cross entropy loss. Experimental results on KITTI2012 and KITTI2015 demonstrate that the proposed method can improve the operation speed while maintaining high accuracy.