Convolution Stage Research Articles

Convolutional Autoencoders for Signal Reconstruction and their Application to Damage Signature Extraction

The utilization of Machine Learning (ML) techniques for Structural Health Monitoring (SHM) has increased in the last years. Using ML techniques for signal reconstruction has been explored in the literature of this field. In this work, a database of Ultrasonic Guided Waves (UGW), propagated through a thermoplastic composite plate, and generated and acquired with piezoelectric transducers (PZT), has been preprocessed, organized, and analyzed to reconstruct signals by using Convolutional Autoencoders (cAE) neural networks. This special type of autoencoders (AE) is based on the utilization of convolutional stages, which are formed by deep-learning layers that are especially useful for their utilization on discretized time-series signals, due to their ability to convolve different sections of the signal, thus extracting different features which are automatically learned by the network. The cAE, which have a bottleneck section in their middle part (between the encoder and the decoder sections), have the ability to reduce the dimensionality of the input data, thus allowing the extraction of a minimum number of relevant features of the signals. Summarizing, this paper presents a neural network designed as a cAE which is able to reconstruct data from a reduced number of features, reaching a correlation value between the real and the artificial data higher than 98%.

MMAIndoor: Patched MLP and multi-dimensional cross attention based self-supervised indoor depth estimation

Depth estimation can provide auxiliary information for scene perception. Generally, extensive textureless surfaces, such as walls and ceilings, exist in indoor environments, and they share similar scene and semantic content. Overly consistent features of local textureless areas fail to reflect changes in depth information, thus degrading the performance of existing depth estimation methods. In response to this challenge, we propose a special indoor depth estimation method, named as MMAIndoor, which can provide global semantic guidance and shape priors for local textureless depth estimation. The depth estimation network is designed efficiently, encompassing the initial convolutional stage and the latent patched multi-layer perceptron (Pat-MLP) stage. The novel Pat-MLP block utilizes MLP partitioning to globally model depth-local information from the convolutional stage and it incorporate axial shift operations to extract local information from various spatial locations, suppressing the smoothing effect of MLP and improving precise estimation of sharp depth changes or small structures indoors. Further, we build a multi-dimensional cross attention (MCA) module to address the weak correlation of the current residual connections for the global context. This MCA captures global dependencies across multi-dimensions by sequentially executing cross attention on both channels and spatial, and effectively mitigate semantic gaps in residual connections. Sufficient experimental results demonstrate the state-of-the-art performance of MMAIndoor on benchmark datasets including NYUv2, ScanNet, and InteriorNet.

Depth Completion with Anisotropic Metric, Convolutional Stages, and Infinity Laplacian

Depth map estimation is crucial for a wide range of applications. Unfortunately, it often presents missing or unreliable data. The objective of depth completion is to fill in the “holes” in a depth map by propagating the depth information using guidance from other sources of information, such as color. Nowadays, classical image processing methods have been outperformed by deep learning techniques. Nevertheless, these approaches require a significantly large number of images and enormous computing power for training. This fact limits their usability and makes them not the best solution in some resource-constrained environments. Therefore, this paper investigates three simple hybrid models for depth completion. We explore a hybrid pipeline that combines a very efficient and powerful interpolator (infinity Laplacian or AMLE) and a series of convolutional stages. The contributions of this article are (i) the use a Texture+Structuredecomposition as a pre-filter stage; (ii) an objective evaluation with three different approaches using KITTI and NYU_V2 data sets; (iii) the use of an anisotropic metric as a mechanism to improve interpolation; and iv) the inclusion of an ablation test. The main conclusions of this work are that using an anisotropic metric improves model performance, and the ablation test demonstrates that the model’s final stage is a critical component in the pipeline; its suppression leads to an approximate 4% increase in MSE. We also show that our model outperforms state-of-the-art alternatives with similar levels of complexity.

The evolution of object detection methods

Object detection is one of the most important domains in computer vision tasks, which is an important branch of artificial intelligence. It aims at finding and locating the accurate position of objects in given pictures or videos. With the development of deep learning techniques, more powerful and robust algorithms have emerged to deal with multi-scale, high-level features to overcome the limitations of traditional pipeline of object detectors. The popularity of transformer framework enables larger capacity datasets by processing self-attention mechanism, and the object detection methods have evolved into a new era. This paper first reviews traditional object detection pipeline and brief history of deep learning, afterwards it focuses on the classification of deep learning-based object detection methods covering Convolution Neural Network based and transformer-based methods. Commonly used datasets and metrics are also covered in the next part. The Convolution Neural Network based methods mainly contain two-stage and one-stage detectors, Convolution Neural Network is the underlying structure of these methods convolutional stages are fundamental parts. Transformer-based models convert traditional object detection issues into end-to-end detection, which is widely used in dealing with images. Finally, the promising future of object detection areas are listed to show guidance on future work.

Pest recognition based on multi-image feature localization and adaptive filtering fusion.

Accurate recognition of pest categories is crucial for effective pest control. Due to issues such as the large variation in pest appearance, low data quality, and complex real-world environments, pest recognition poses challenges in practical applications. At present, many models have made great efforts on the real scene dataset IP102, but the highest recognition accuracy is only 75%. To improve pest recognition in practice, this paper proposes a multi-image fusion recognition method. Considering that farmers have easy access to data, the method performs fusion recognition on multiple images of the same pest instead of the conventional single image. Specifically, the method first uses convolutional neural network (CNN) to extract feature maps from these images. Then, an effective feature localization module (EFLM) captures the feature maps outputted by all blocks of the last convolutional stage of the CNN, marks the regions with large activation values as pest locations, and then integrates and crops them to obtain the localized features. Next, the adaptive filtering fusion module (AFFM) learns gate masks and selection masks for these features to eliminate interference from useless information, and uses the attention mechanism to select beneficial features for fusion. Finally, the classifier categorizes the fused features and the soft voting (SV) module integrates these results to obtain the final pest category. The principle of the model is activation value localization, feature filtering and fusion, and voting integration. The experimental results indicate that the proposed method can train high-performance feature extractors and classifiers, achieving recognition accuracy of 73.9%, 99.8%, and 99.7% on IP102, D0, and ETP, respectively, surpassing most single models. The results also show that thanks to the positive role of each module, the accuracy of multi-image fusion recognition reaches the state-of-the-art level of 96.1%, 100%, and 100% on IP102, D0, and ETP using 5, 2, and 2 images, respectively, which meets the requirements of practical applications. Additionally, we have developed a web application that applies our research findings in practice to assist farmers in reliable pest identification and drive the advancement of smart agriculture.

Transformer-based rapid human pose estimation network

Most current human pose estimation methods pursue excellent performance via large models and intensive computational requirements, resulting in slower models. These methods cannot be effectively adopted for human pose estimation in real applications due to their high memory and computational costs. To achieve a trade-off between accuracy and efficiency, we propose TRPose, a Transformer-based network for human pose estimation rapidly. TRPose consists of an early convolutional stage and a later Transformer stage seamlessly. Concretely, the convolutional stage forms a Rapid Fusion Module (RFM), which efficiently acquires multi-scale features via three parallel convolution branches. The Transformer stage utilizes multi-resolution Transformers to construct a Dual scale Encoder Module (DEM), aiming at learning long-range dependencies from different scale features of the whole human skeletal keypoints. The experiments show that TRPose acquires 74.3 AP and 73.8 AP on COCO validation and testdev datasets with 170+ FPS on a GTX2080Ti, which achieves the better efficiency and effectiveness trade-offs than most state-of-the-art methods. Our model also outperforms mainstream Transformer-based architectures on MPII dataset, yielding 89.9 PCK@0.5 score on val set without extra data.

Improved Tomographic Estimates by Specialized Neural Networks

AbstractCharacterization of quantum objects, being states, processes, or measurements, complemented by previous knowledge about them is a valuable approach, especially as it leads to routine procedures for real‐life components. To this end, machine learning algorithms have demonstrated to successfully operate in presence of noise, especially for estimating specific physical parameters. Here, it is shown that a neural network (NN) can improve the tomographic estimate of parameters by including a convolutional stage. This technique is applied to quantum process tomography for the characterization of several quantum channels. A stable and reliable operation is demonstrated that is achievable by training the network only with simulated data. The obtained results show the viability of this approach as an effective tool based on a completely new paradigm for the employment of NNs operating on classical data produced by quantum systems.

Human Activity Recognition Using Attention-Mechanism-Based Deep Learning Feature Combination.

Human activity recognition (HAR) performs a vital function in various fields, including healthcare, rehabilitation, elder care, and monitoring. Researchers are using mobile sensor data (i.e., accelerometer, gyroscope) by adapting various machine learning (ML) or deep learning (DL) networks. The advent of DL has enabled automatic high-level feature extraction, which has been effectively leveraged to optimize the performance of HAR systems. In addition, the application of deep-learning techniques has demonstrated success in sensor-based HAR across diverse domains. In this study, a novel methodology for HAR was introduced, which utilizes convolutional neural networks (CNNs). The proposed approach combines features from multiple convolutional stages to generate a more comprehensive feature representation, and an attention mechanism was incorporated to extract more refined features, further enhancing the accuracy of the model. The novelty of this study lies in the integration of feature combinations from multiple stages as well as in proposing a generalized model structure with CBAM modules. This leads to a more informative and effective feature extraction technique by feeding the model with more information in every block operation. This research used spectrograms of the raw signals instead of extracting hand-crafted features through intricate signal processing techniques. The developed model has been assessed on three datasets, including KU-HAR, UCI-HAR, and WISDM datasets. The experimental findings showed that the classification accuracies of the suggested technique on the KU-HAR, UCI-HAR, and WISDM datasets were 96.86%, 93.48%, and 93.89%, respectively. The other evaluation criteria also demonstrate that the proposed methodology is comprehensive and competent compared to previous works.

Counting piglet suckling events using deep learning-based action density estimation

Analysis of piglet suckling behaviour is important for the evaluation of piglet nutrient ingestion, health, welfare, and affinity with the sow. In this study, an action density estimation network (ADEN) was proposed for counting the events of piglet suckling followed by automated analysis of suckling behaviour. ADEN is a two-stream network primarily composed of 1) a network stream that processes video images with a higher frame rate (faster stream) and 2) a network stream that processes video images with a lower frame rate (slower stream). Each stream consists of a ResNet-50 with five convolutional stages. A multi-stage attention connection (MSAC), composed of four Spatial-Temporal-Channel (STC) multi-attention structures, is proposed to bridge Faster Stream and Slower Stream and capture discriminative features. The output attention features from each faster stream stage are laterally fused into the corresponding slower stream stage in a concatenating manner. Following this, the features from the last convolutional stages in the Slow stream and Fast stream are fused using concatenation and are then decoded by using three convolutional layers. The last convolutional layer outputs a heatmap on the action density of piglet suckling behaviour. Finally, the number of suckling events is predicted by integrating all the pixel values in the heatmap. Experimental and comparative tests were conducted to validate the effectiveness of the proposed ADEN with a training dataset and a test dataset from 14 pig pens. The 507 video clips (126,750 images for 7 h) from the 1-9th pens were selected as training datasets. The 143 video clips (35,750 images for 2 h) from the 10-13th pens were selected as short-term test datasets. One untrimmed video (162,000 images for 9 h) from the 14th pen was used to ultimately evaluate the action density estimation performance of the ADEN. ADEN was compared with seven approaches and its superiority was demonstrated with an r = 0.938, an RMSE = 1.080, and a MAE = 0.967 in short video clips and r = 0.982, MAE = 0.161, and RMSE = 0.563 in the untrimmed long video. The ADEN proved it feasible to predict the number of suckling events by using action density estimation.

EfferDeepNet: An Efficient Semantic Segmentation Method for Outdoor Terrain

The recognition of terrain and outdoor complex environments based on vision sensors is a key technology in practical robotics applications, and forms the basis of autonomous navigation and motion planning. While traditional machine learning methods can be applied to outdoor terrain recognition, their recognition accuracy is low. In order to improve the accuracy of outdoor terrain recognition, methods based on deep learning are widely used. However, the network structure of deep learning methods is very complex, and the number of parameters is large, which cannot meet the actual operating requirements of of unmanned systems. Therefore, in order to solve the problems of poor real-time performance and low accuracy of deep learning algorithms for terrain recognition, this paper proposes the efficient EfferDeepNet network for pixel level terrain recognition in order to realize global perception of outdoor environment. First, this method uses convolution kernels with different sizes in the depthwise separable convolution (DSC) stage to extract more semantic feature information. Then, an attention mechanism is introduced to weight the acquired features, focusing on the key local feature areas. Finally, in order to avoid redundancy due to a large number of features and parameters in the model, this method uses a ghost module to make the network more lightweight. In addition, to solve the problem of pixel level terrain recognition having a negative effect on image boundary segmentation, the proposed method integrates an enhanced feature extraction network. Experimental results show that the proposed EfferDeepNet network can quickly and accurately perform global recognition and semantic segmentation of terrain in complex environments.

AGD-Linknet: A Road Semantic Segmentation Model for High Resolution Remote Sensing Images Integrating Attention Mechanism, Gated Decoding Block and Dilated Convolution

Road information is a kind of important geographic information. Road information extracted from remote sensing images is widely used in map, traffic, navigation, military and many other fields. However, the autonomous extraction of road information from high resolution remote sensing images has some problems such as incoherence, incompleteness and poor connectivity, therefore, a semantic segmentation model for roads in high resolution remote sensing images, called AGD-Linknet, is proposed, which integrates attention mechanisms, gated decoder block, and dilated convolution. This model mainly consists of three parts.Firstly, the stem block is used as the initial convolution layer of the model to reduce the information loss in the convolution stage; Secondly, the series-parallel combined dilated convolution and coordinate attention block into the center of the network, which enlarges the receptive field of the network and improves the feature extraction ability of spatial domain and channel domain information; Finally, in the decoder part, gated convolution is introduced in the decoder part to improve the extraction of road edge. Compared with U-Net, Linknet and D-Linknet on the DeepGlobe dataset, the proposed AGD-Linknet has improved the pixel accuracy, mean intersection over union and F1-Score index of road recognition by 1.41%-11.52%, 0.0077-0.1473, 0.0057-0.1292, and has certain effectiveness and feasibility in many scenarios in rural areas, urban, and suburbs. And can be apply to the tasks of road recognition and extraction in high-resolution remote sensing.

Method of convolutional neural network with hybrid attention for crack detection

In recent years, automatic image crack detection has become a critical task for ensuring the safety of various facilities. Many researchers utilized convolutional neural networks (CNNs) for crack detection. However, the existing CNN methods have the limitation of the receptive fields and struggle to establish long dependencies and global background information. Aiming to address this problem, we propose a CNN method with an attention module and recurrent mechanism. In terms of the backbone network, we propose a backbone network named Attention-VGG-19. In our network, we remove the last three fully connected layers of VGG-19 to expect the meaningful side output with different scales and to reduce the number of parameters. The fully connected layers are computationally intensive. The proposed network also provides integrated direct supervision for features of each convolutional stage. Then, to promote important features and suppress unimportant features, an attention module called CrackP attention is proposed. In the recurrent mechanism, the module can be unrolled into R loops to connect the path between one pixel and its neighboring ones. To prove the effectiveness of our proposed method, we evaluate it on four public crack datasets, DeepCrack, crack forest dataset (CFD), Crack500, and CrackTree260, on which it achieves F-score (F1) values of 0.876, 0.652, 0.692, and 0.593, respectively. After a guided filter is added, our method achieves F1 values of 0.878, 0.638, 0.681, and 0.526 on these four datasets.

An effective industrial defect classification method under the few-shot setting via two-stream training

Visual surface defect inspection provides an important tool for product quality assessment in a wide range of industrial applications. In recent years, diverse convolutional neural network (CNN) models have been developed for the high-accuracy image classification of natural objects. However, it remains a challenging task to develop deep-learning-based approaches for surface defect inspection of industrial products. The main challenge is that surface defects of industrial products occur in low probability, therefore it is impractical to build a large image dataset containing different types of defects to train/fine-tune deep CNN models to achieve satisfactory performance when applied to industrial inspection tasks. To overcome the above-mentioned problem, we build a light-weight defect classification model based on the pre-trained SqueezeNet architecture and present three effective techniques to achieve high-accuracy defect classification using very few collected defective training samples. First, we experimentally demonstrate that it is reasonable to freeze shallower convolutional layers and only fine-tune deeper layers in the last convolutional stage when there are not enough training samples. Second, we integrate batch-size independent Group Normalization (GN) into SqueezeNet to compute stable statistics based on a limited number of defective samples for training the classification model. Third, we add an auxiliary task to perform the similarity comparison of two defective samples during the training phase, significantly increasing the size of the training data and generating more distinctive defect-specific representations. The experimental results on the NEU-CLS dataset and USB-FS dataset verify the effectiveness of the proposed method. Our method can obtain accurate defect classification results when very few defective training samples are provided, i.e., about 97.69% and 82.92% average classification accuracy on the NEU-CLS and USB-FS datasets, respectively, under the 5-shot setting (5 images per defect category).

Single image super‐resolution based on progressive fusion of orientation‐aware features

Single image super-resolution (SISR) is an active research topic in the fields of image processing, computer vision and pattern recognition, restoring high-frequency details and textures based on the low-resolution input image. In this paper, we aim to build more accurate and faster SISR models via developing better-performing feature extraction and fusion techniques. Firstly, we proposed a novel Orientation-Aware feature extraction/selection Module (OAM), which contains a mixture of 1D and 2D convolutional kernels (i.e., 3×1, 1×3, and 3×3) for extracting orientation-aware features. The channel attention mechanism is deployed within each OAM, performing scene-specific selection of informative outputs of the orientation-dependent kernels (e.g., horizontal, vertical, and diagonal). Secondly, we present an effective fusion architecture to progressively integrate multi-scale features extracted in different convolutional stages. Instead of directly combining low-level and high-level features, similar outputs of adjacent feature extraction modules are grouped and further compressed to generate a more concise representation of a specific convolutional stage for high-accuracy SISR task. Based on the above two important improvements, we present a compact but effective CNN-based model for high-quality SISR via Progressive Fusion of Orientation-Aware features (SISR-PF-OA). Extensive experimental results verify the superiority of the proposed SISR-PF-OA model, performing favorably against the state-of-the-art models in terms of both restoration accuracy and computational efficiency (e.g., SISR-PF-OA outperforms RCAN model, achieving higher PSNR 31.25 dB vs. 31.21 dB and using fewer FLOPs 764.41 G vs. 1020.28 G on the Manga109 dataset for scale factor ×4 SISR task.). The source codes will be made publicly available.

Method of feature pyramid and attention enhancement network for pavement crack detection

The detection of pavement crack information plays an important role in the service evaluation of infrastructure engineering performance. The randomness of road noise, pollution, and spots make it difficult to ensure the accuracy of traditional image processing algorithms to detect cracks. Therefore, the crack identification technology under complex background conditions is particularly important. We propose a network architecture, named feature pyramid and attention enhancement network (FPAEN) for pavement crack detection. Through a new multiscale feature fusion module, the deep semantic information is integrated into the low-level convolution stage layer by layer, so the context information circulates throughout the whole network. In addition, the attention enhancement mechanism and the multiscale dilated convolution module are used to strengthen the network model’s ability to locate the crack pixels. To evaluate the performance of the proposed approach, we train FPAEN on the DeepCrack dataset and test it on four datasets. Experimental results show that our method outperforms other advanced crack detection approaches in F-score and mean intersection over union.

VC-YOLO: Towards Real-Time Object Detection in Aerial Images

Object detection for aerial images is a crucial and challenging task in the field of computer vision. Previous CNN-based methods face problems related to extreme variation of object scales and the complex background in aerial images, which vary significantly from natural scenes. On the other hand, a great many of existing detectors highly rely on computational performance and cannot handle real-time tasks. To address this problems, we propose a lightweight real-time object detection network which is named VC-YOLO. In the backbone part, we introduce a receptive field extended backbone with limited number of convolution layers to learn the features and context information of various objects. In the detection part, channel attention module and spatial attention module are used to generate discriminative feature representation. To make full use of semantic feature maps in backbone network, we improve the feature pyramid network (FPN) with more lateral connections to reuse the features in each convolution stage. We evaluate VC-YOLO on NWPU VHR-10 and VisDrone benchmark datasets. Experimental results show that VC-YOLO achieves superior detection accuracy with high efficiency compared with the existing methods.

PrimeNet: Adaptive Multi-Layer Deep Neural Structure for Enhanced Feature Selection in Early Convolution Stage

The colossal depths of the deep neural network sometimes suffer from ineffective backpropagation of the gradients through all its depths, whereas the strong performance of shallower multilayer neural structures proves their ability to increase the gradient signals in the early stages of training, which easily gets backpropagated for global loss corrections. Shallow neural structures are always a good starting point for encouraging the sturdy feature characteristics of the input. In this research, a shallow, deep neural structure called PrimeNet is proposed. PrimeNet is aimed to dynamically identify and encourage the quality visual indicators from the input to be used by the subsequent deep network layers and increase the gradient signals in the lower stages of the training pipeline. In addition to this, the layer-wise training is performed with the help of locally generated errors, which means the gradient is not backpropagated to previous layers, and the hidden layer weights are updated during the forward pass, making this structure a backpropagation free variant. PrimeNet has obtained state-of-the-art results on various image datasets, attaining the dual objective of: (1) a compact dynamic deep neural structure, which (2) eliminates the problem of backward-locking. The PrimeNet unit is proposed as an alternative to traditional convolution and dense blocks for faster and memory-efficient training, outperforming previously reported results aimed at adaptive methods for parallel and multilayer deep neural systems.

Optimization of Artistic Image Segmentation Algorithm Based on Feed Forward Neural Network under Complex Background Environment.

Based on the theory and application, this paper discusses the optimization of art image segmentation algorithm based on FFNN (Feed Forward Neural Network). In this paper, residual units are used in the corresponding stages of encoder and decoder, and feature information of several convolution layers in each convolution stage of encoder is extracted at the same time. And the feature pyramid module is used to extract multiscale features from the feature map of the last convolution stage in the encoder. Finally, pixel by pixel additions combine the previously mentioned feature information into the corresponding layer of the decoder. Additionally, an improved weight adaptive algorithm based on feature preservation is suggested in this paper, which addresses the issue that the conventional image segmentation algorithm is noise-sensitive. The adaptive connection weight mechanism is also introduced. The accuracy and recall rates of this optimization algorithm can both reach 96.574%, according to the results of 50% cross-validation. All the segmentation performance evaluation indexes of this algorithm are higher than the existing main algorithms. Moreover, the algorithm takes a short time, does not need too much manual intervention, and can effectively segment artistic images. The optimization algorithm in this paper has certain reference significance for the related research of artistic image segmentation.

An ultra-fast bi-phase advanced network for segmenting crop plants from dense weeds

Fast and accurate perception of crop plants is critical for robotic weeding. However, due to the similar appearance, it remains challenging to precisely and efficiently segment crop plants from weeds, especially when the weed density is high. To tackle this problem, we devise an ultra-fast bi-phase advanced network (BA-Net) architecture that advances both the feature extraction phase and the model training phase. A bi-path fusion tree structure and a bi-polar distributed loss are introduced for the two phases, respectively. The bi-path fusion tree merges features from two adjacent convolutional stages at each step and efficiently models the scene through a tree-like structure. The bi-polar distributed loss strengthens the training at ambiguous areas, which can greatly enhance the distinctness at detailed structures like the crop plant boundaries. Experiments demonstrate the superiority of the proposed BA-Net to other state-of-the-art methods in terms of both accuracy and efficiency. BA-Net can accurately segment the crop plants from the weedy background, achieving an intersection over union score of 0.963, a F-measure score of 0.981, and a mean absolute error of 0.00376. Also, BA-Net is lightweight with only 0.55 M trainable parameters and achieves ultra-fast speed, i.e. , over 340 FPS on a desktop and over 26 FPS on a Jetson TX2 embedded computer. The code of BA-Net is available at https://github.com/ZhangXG001/BA-Net . • An ultra-fast deep network is proposed for segmenting crop plants from dense weeds. • The bi-path fusion tree structure can perform efficient multi-scale feature fusion. • The bi-polar distributed loss can improve the segmentation at the obscure areas. • The network achieves superior accuracy and efficiency over state-of-the-arts. • The network runs over 26 FPS on a Jetson TX2 embedded computer.

Multi-attention augmented network for single image super-resolution

How to improve the representational power of visual features extracted by deep convolutional neural networks is of crucial importance for high-quality image super-resolution. To address this issue, we propose a multi-attention augmented network, which mainly consists of content-, orientation- and position-aware modules. Specifically, we develop an attention augmented U-net structure to form the content-aware module in order to learn and combine multi-scale informative features within a large receptive field. To better reconstruct image details in different directions, we design a set of pre-defined sparse kernels to construct the orientation-aware module, which can extract more representative multi-orientation features and enhance the discriminative capacity in stacked convolutional stages. Then these extracted features are adaptively fused through channel attention mechanism. In upscale stage, the position-aware module adopts a novel self-attention to reweight the element-wise value of final low-resolution feature maps, for further suppressing the possible artifacts. Experimental results demonstrate that our method obtains better reconstruction accuracy and perceptual quality against state-of-the-art methods.