Cross-layer Connection Research Articles

SWANN: Small-World Architecture for Fast Convergence of Neural Networks

On-device intelligence has become increasingly widespread in the modern smart application landscape. A standing challenge for the applicability of on-device intelligence is the excessively high computation cost of training highly accurate Deep Learning (DL) models. These models require a large number of training iterations to reach a high convergence accuracy, hindering their applicability to resource-constrained embedded devices. This paper proposes a novel transformation which changes the topology of the DL architecture to reach an optimal cross-layer connectivity. This, in turn, significantly reduces the number of training iterations required for reaching a target accuracy. Our transformation leverages the important observation that for a set level of accuracy, convergence is fastest when network topology reaches the boundary of a Small-World Network. Small-world graphs are known to possess a specific connectivity structure that enables enhanced signal propagation among nodes. Our small-world models, called SWANNs, provide several intriguing benefits: they facilitate data (gradient) flow within the network, enable feature-map reuse by adding long-range connections and accommodate various network architectures/datasets. Compared to densely connected networks (e.g., DenseNets), SWANNs require a substantially fewer number of training parameters while maintaining a similar level of classification accuracy. We evaluate our networks on various DL model architectures and image classification datasets, namely, MNIST, CIFAR10, CIFAR100, and ImageNet. Our experiments demonstrate an average of ≈2.1× improvement in convergence speed to the desired accuracy.

Craniofacial Reconstruction via Face Elevation Map Estimation Based on the Deep Convolution Neutral Network

In this study, to achieve the possibility of predicting face by skull automatically, we propose a craniofacial reconstruction method based on the end-to-end deep convolutional neural network. Three-dimensional volume data are obtained from 1447 head CT scans of Chinese people of different ages. The facial and skull surface data are projected onto two-dimensional space to generate a two-dimensional elevation map, and then, use the deep convolution neural network to realize the prediction of skull to face shape in two-dimensional space. The encoder and decoder are composed of first feature extraction through the encoder and then as the input of the decoder to generate the craniofacial restoration image. In order to accurately describe the features of different scales, we adopt an U-shaped codec structure with cross-layer connections. Therefore, the output features are decomposed with the features of the corresponding scales in the encoding stage to achieve the integration of different scales while restoring the feature scales in the compression and decoding stage. Meanwhile, the U-net structures help to avoid the problem of loss of detail features in the downsampling process. We use supervised learning to obtain the prediction model from skull to facial elevation map. Back-projection operation is performed afterwards to generate facial surface data in 3D space. Experiments show that the proposed method in this study can effectively achieve craniofacial reconstruction, and for most part of the face, restoration error is controlled within 2 mm.

From WASD to BLS with application to pattern classification

The single-hidden-layer feedforward neural network (SLFN) exhibits an excellent approximation ability, despite its simple structure. This study examines two neural network models developed based on the general structure of the SLFN, i.e., the weights-and-structure-determination (WASD) neural network and broad learning system (BLS). The BLS and WASD neural networks exhibit many similarities in terms of their structure and algorithm, while BLS incorporates certain distinct ideas that render it an improved version of the WASD neural network. The research described in this paper is the first approach to establish a connection between the BLS and WASD neural networks. Moreover, pattern classification experiments on a foot dataset and several benchmark datasets are conducted using these two kinds of neural network models, and the performance of different models is compared and analyzed.Experimental results show that dividing the feature nodes into one group is the best choice for our foot dataset, and that the cross-layer connectivity and sparse coding of BLS can effectively improve the classification accuracy. In addition, it is shown that the dynamic update algorithm of BLS can improve the accuracy and save a lot of time when updating the model structure, but the final accuracy depends on the initial structure and parameters. Comparative experiments show that BLS achieves 89.12% classification accuracy for the foot dataset, which is the highest among all the demonstrated WASD neural networks and two other SLFN models. Finally, extended experiments on five benchmark datasets show that BLS can achieve satisfactory accuracy on multiple types of datasets, while WASD neural networks are possibly inadequate in processing image datasets in contrast.

DefogNet: A Single-Image Dehazing Algorithm with Cyclic Structure and Cross-Layer Connections

Inspired by the application of CycleGAN networks to the image style conversion problem Zhu et al. (2017), this paper proposes an end-to-end network, DefogNet, for solving the single-image dehazing problem, treating the image dehazing problem as a style conversion problem from a fogged image to a nonfogged image, without the need to estimate a priori information from an atmospheric scattering model. DefogNet improves on CycleGAN by adding a cross-layer connection structure in the generator to enhance the network’s multiscale feature extraction capability. The loss function was redesigned to add detail perception loss and color perception loss to improve the quality of texture information recovery and produce better fog-free images. In this paper, the novel Defog-SN algorithm is presented. This algorithm adds a spectral normalization layer to the discriminator’s convolution layer to make the discriminant network conform to a 1-Lipschitz continuum and further improve the model’s stability. In this study, the experimental process is completed based on the O-HAZE, I-HAZE, and RESIDE datasets. The dehazing results show that the method outperforms traditional methods in terms of PSNR and SSIM on synthetic datasets and Avegrad and Entropy on naturalistic images.

A Multi-Task CNN for Maritime Target Detection

In this letter, we construct MaRine ShiP (MRSP-13), a novel dataset containing 37,161 ship target images belonging to 13 classes with bounding box annotation, and among them there are 3051 images labeled with pixel-level annotation. This dataset equips us with the capability to conduct baseline experiments on maritime target classification, detection and segmentation. We propose a cross-layer multi-task CNN model for maritime target detection, which can simultaneously solve ship target detection, classification, and segmentation. Experimental results have demonstrated the efficiency of the MRSP-13 dataset to be used for maritime target analysis. In addition, the results validate the fact that by adopting the strategies of feature sharing, joint learning, and cross-layer connections, the proposed model achieves superior performance with less annotations. We believe that our MRSP-13 dataset and corresponding baseline experiments will lay down the foundation for further research in maritime target processing.

Research on a New Convolutional Neural Network Model Combined With Random Edges Adding

It is always a hot and difficult point to improve the accuracy of the convolutional neural network model and speed up its convergence. Based on the idea of the small world network, a random edge adding algorithm is proposed to improve the performance of the convolutional neural network model. This algorithm takes the convolutional neural network model as a benchmark and randomizes backwards and cross layer connections with probability p to form a new convolutional neural network model. The proposed idea can optimize the cross-layer connectivity by changing the topological structure of the convolutional neural network and provide a new idea for the improvement of the model. The simulation results based on Fashion-MINST and cifar10 data set show that the model recognition accuracy and training convergence speed are greatly improved by random edge adding reconstructed models with a probability of p = 0.1.

The face image super-resolution algorithm based on combined representation learning

Face super-resolution reconstruction is the process of predicting high-resolution face images from one or more observed low-resolution face images, which is a typical pathological problem. As a domain-specific super-resolution task, we can use facial priori knowledge to improve the effect of super-resolution. We propose a method of face image super-resolution reconstruction based on combined representation learning method, using deep residual networks and deep neural networks as generators and discriminators, respectively. First, the model uses residual learning and symmetrical cross-layer connection to extract multilevel features. Local residual mapping improves the expressive capability of the network to enhance performance, solves gradient dissipation in network training, and reduces the number of convolution cores in the model through feature reuse. The feature expression of the face image at the high-dimensional visual level is obtained. The visual feature is sent to the decoder through the cross-layer connection structure. The deconvolution layer is used to restore the spatial dimension gradually and repair the details and texture features of the face. Finally, combine the attention block and the residual block reconstruction in the deep residual network to super-resolution face images that are highly similar to high-resolution images and difficult to be discriminated by the discriminator. On this basis, combined representation learning is conducted to obtain numerous realistic results of visual perception. The experimental results on the face datasets can show that the Peak Signal-to-Noise Ratio of the proposed method is improved.

Multi-scale traffic vehicle detection based on faster R–CNN with NAS optimization and feature enrichment

It well known that vehicle detection is an important component of the field of object detection. However, the environment of vehicle detection is particularly sophisticated in practical processes. It is comparatively difficult to detect vehicles of various scales in traffic scene images, because the vehicles partially obscured by green belts, roadblocks or other vehicles, as well as influence of some low illumination weather. In this paper, we present a model based on Faster R–CNN with NAS optimization and feature enrichment to realize the effective detection of multi-scale vehicle targets in traffic scenes. First, we proposed a Retinex-based image adaptive correction algorithm (RIAC) to enhance the traffic images in the dataset to reduce the influence of shadow and illumination, and improve the image quality. Second, in order to improve the feature expression of the backbone network, we conducted Neural Architecture Search (NAS) on the backbone network used for feature extraction of Faster R–CNN to generate the optimal cross-layer connection to extract multi-layer features more effectively. Third, we used the object Feature Enrichment that combines the multi-layer feature information and the context information of the last layer after cross-layer connection to enrich the information of vehicle targets, and improve the robustness of the model for challenging targets such as small scale and severe occlusion. In the implementation of the model, K-means clustering algorithm was used to select the suitable anchor size for our dataset to improve the convergence speed of the model. Our model has been trained and tested on the UN-DETRAC dataset, and the obtained results indicate that our method has art-of-state detection performance.

A Deep Convolutional Network for Saliency Object Detection with Balanced Accuracy and High Efficiency

It is difficult for current salient object detection algorithms to reach a good balance performance between accuracy and efficiency. To solve this problem, a deep convolutional network for saliency object detection with balanced accuracy and high efficiency is produced. First, through replacing the traditional convolution with the decomposed convolution, the computational complexity is greatly reduced and the detection efficiency of the model is improved. Second, in order to make better use of the characteristics of different scales, sparse cross-layer connection structure and multi-scale fusion structure are adopted to improve the detection precision. A wide range of evaluations show that compared with the existing methods, the proposed algorithm achieves the leading performance in efficiency and accuracy.

Medical Image Segmentation Algorithm Based on Optimized Convolutional Neural Network-Adaptive Dropout Depth Calculation

Medical image segmentation is a key technology for image guidance. Therefore, the advantages and disadvantages of image segmentation play an important role in image-guided surgery. Traditional machine learning methods have achieved certain beneficial effects in medical image segmentation, but they have problems such as low classification accuracy and poor robustness. Deep learning theory has good generalizability and feature extraction ability, which provides a new idea for solving medical image segmentation problems. However, deep learning has problems in terms of its application to medical image segmentation: one is that the deep learning network structure cannot be constructed according to medical image characteristics; the other is that the generalizability y of the deep learning model is weak. To address these issues, this paper first adapts a neural network to medical image features by adding cross-layer connections to a traditional convolutional neural network. In addition, an optimized convolutional neural network model is established. The optimized convolutional neural network model can segment medical images using the features of two scales simultaneously. At the same time, to solve the generalizability problem of the deep learning model, an adaptive distribution function is designed according to the position of the hidden layer, and then the activation probability of each layer of neurons is set. This enhances the generalizability of the dropout model, and an adaptive dropout model is proposed. This model better addresses the problem of the weak generalizability of deep learning models. Based on the above ideas, this paper proposes a medical image segmentation algorithm based on an optimized convolutional neural network with adaptive dropout depth calculation. An ultrasonic tomographic image and lumbar CT medical image were separately segmented by the method of this paper. The experimental results show that not only are the segmentation effects of the proposed method improved compared with those of the traditional machine learning and other deep learning methods but also the method has a high adaptive segmentation ability for various medical images. The research work in this paper provides a new perspective for research on medical image segmentation.

A realized on-demand cross-layer connection strategy for wireless self-organization link based on WLAN

For massive data communication, the wireless network is often needed to support data transmission in complex environment. In this case, a portable, inexpensive and efficient wireless device is especially helpful. In order to realize this function, we propose a wireless self-organized connection strategy in this paper. We focus on a multihop wireless network with single relay at each hop. One of the challenges over this system is to set up the communication link optimally and automatically, which is also the main contribution of this paper. So we designed an On-Demand Cross-Layer Connection Strategy (ODCLCS) based on WLAN to achieve the data transmission between source and destination point. This strategy crosses application layer, transport layer, network layer and MAC layer. By using active scanning in MAC layer, proper nodes can be found and connected. The automatic routing method can get benefit from active scanning message, automatically allocate IP address for each node. At last, we develop a link maintenance function which works after the link is established. The link maintenance function using heartbeats communication mechanism to find link errors. Moreover, the link maintenance message is very short which can be neglected compared with the data stream. The ODCLCS is designed for Wireless Self-organized Link (WSOL) system. WSOL system aims to establish a stable and user-friendly communication link by using ODCLCS and other necessary programs. All programs are properly written and debugged repeatedly by us, which can work well on testbed devices. Performance evaluation using NS-3 simulator shows the characteristic of WSOL. The WSOL program for testbed is based on cfg80211 and mac80211 in Linux Kernel. And some API based on libnl library is invoked. The testing data provided verify the good performance of ODCLCS in reality.

Pattern Recognition of Partial Discharge Based on VMD-CWD Spectrum and Optimized CNN With Cross-Layer Feature Fusion

In order to improve the recognition accuracy of partial discharge (PD) by making full use of the time-frequency characteristics of PD signals and employing deep learning theory, a kind of PD pattern recognition method based on variational mode decomposition (VMD)-Choi-Williams distribution (CWD) spectrum and optimized convolutional neural network (CNN) with cross-layer feature fusion is proposed in this paper. Firstly, a PD signal is decomposed into several components by VMD algorithm, and the CWD analysis of the obtained components is carried out to obtain the VMD-CWD time-frequency spectrum. Secondly, the cross-layer feature fusion and optimization CNN (CFFO-CNN) is constructed by introducing cross-layer connection and optimization algorithm. Thirdly, the VMD-CWD is regarded as the input vector to train CFFO-CNN to learn and extract the intrinsic features of the spectrum. Finally, the trained network is used to recognize the PD types of the testing samples. The proposed method is compared with traditional recognition methods such as BP neural network (BPNN) and support vector machine (SVM), as well as some commonly used deep learning algorithms. The experimental results indicate that the recognition performance of the proposed method is significantly better than that of existing recognition methods with accuracy up to 99.5%. It is proved that CFFO-CNN has superior feature extraction ability, which can extract the internal features of the VMD-CWD spectrum independently with higher recognition accuracy and wider application prospect.

Multiview Scene Image Inpainting Based on Conditional Generative Adversarial Networks

With the help of a multiview system, an unmanned vehicle system can better understand the surrounding environment and choose a more accurate and safer path to avoid obstacles. However, due to the interference of the signal or the loss of part of the signal during the acquisition, processing, compression, transmission and decompression of the video image signal, the local area of the image is abnormal, which affects the perception and decision of the system. This article addresses the problems of inaccurate restored images and noise in the restored images by proposing an image restoration method that is applied to a multicamera system. We utilize different perspective images captured by different cameras to assist and constrain the restoration of the damaged image. This method restores the image by combining sample representations and sample distribution models which respectively based on self-encoder reconstruction loss learning and generative adversarial networks. In this method, the infrastructure is a conditional generative adversarial network, the condition is the images that are from the other perspectives, and the generator is a self-encoder structure with cross-layer connection, group convolution and feature map channel exchanged. This method was carried out on a dataset recorded in Zurich using a pair of cameras mounted on a mobile platform. The experimental results demonstrate that the proposed method is superior to the existing methods in terms of mean L1 Loss, mean L2 Loss and the peak signal to noise ratio (PSNR).

Weighted Densely Connected Convolutional Networks for Reinforcement Learning

A weighted densely connected convolution network (W-DenseNet) is proposed for reinforcement learning in this work. The W-DenseNet can maximize the information flow between all layers in the network by cross layer connection, which can reduce the phenomenon of gradient vanishing and degradation, and greatly improves the speed of training convergence. The weight coefficient introduced in W-DenseNet, the current layer received all the previous layers’ feature maps with different initial weights, which can extract feature information of different layers more effectively according to tasks. According to the weight adjusted by learning, the cross-layer connection is pruned to remove the cross-layer connection with smaller weight, so as to reduce the number of cross-layer. In this work, GridWorld and FlappyBird games are used for simulation. The simulation results of deep reinforcement learning based on W-DenseNet are compared with the traditional deep reinforcement learning algorithm and reinforcement learning algorithm based on DenseNet. The simulation results show that the proposed W-DenseNet method can make the results more convergent, reduce the training time, and obtain more stable results.

Convolutional neural network for spectral–spatial classification of hyperspectral images

Hyperspectral images (HSIs) have great potential in military reconnaissance, land use, marine monitoring and many other fields. In recent years, the convolutional neural network (CNN) has been successfully used to classify hyperspectral data and achieved remarkable performance. However, the limited labeled samples of HSI lead to the small sample size problem, which remains the major challenge for CNN-based HSI classification. Besides, most CNN models have large number of parameters needed to be learned, which cause high computational cost. To address the aforementioned two issues, a novel CNN-based HSI classification method is proposed. The proposed classification method has several distinguishing characteristics. First, the proposed method can robustly extract spectral and spatial features of the HSI simultaneously. Second, in the proposed CNN architecture, all convolution layers are 1 × 1 convolution layer except the first one, which can greatly reduce the number of network parameters, thus accelerating the training and testing process. Third, a small convolution and feature reuse (SC-FR) module is developed. The SC-FR module is composed of two composite layers and each composite layer consists of two cascaded 1 × 1 convolution layers. Through cross-layer connecting, the input and output features of each composite layer are concatenated and passed to the next convolution layer, thus achieving feature reuse mechanism. Cross-layer connection increases information flow and the utilization rate of middle-level features, which enhances the generalization performance of CNN effectively. Experimental results on three benchmark HSIs demonstrate the competitive superiority of the proposed method over several state-of-the-art HSI classification methods, especially when training samples are limited.

Underwater Image Restoration Based on a Parallel Convolutional Neural Network

Restoring degraded underwater images is a challenging ill-posed problem. The existing prior-based approaches have limited performance in many situations due to the reliance on handcrafted features. In this paper, we propose an effective convolutional neural network (CNN) for underwater image restoration. The proposed network consists of two paralleled branches: a transmission estimation network (T-network) and a global ambient light estimation network (A-network); in particular, the T-network employs cross-layer connection and multi-scale estimation to prevent halo artifacts and to preserve edge features. The estimates produced by these two branches are leveraged to restore the clear image according to the underwater optical imaging model. Moreover, we develop a new underwater image synthesizing method for building the training datasets, which can simulate images captured in various underwater environments. Experimental results based on synthetic and real images demonstrate that our restored underwater images exhibit more natural color correction and better visibility improvement against several state-of-the-art methods.

Customer Churn Prediction Model Based on LSTM and CNN in Music Streaming

It is critical for companies to accurately predict customer churn if they want to achieve sustained development. Previous studies have used many machine learning methods to predict customer churn. The common models cannot make full use of time series features. To overcome this shortcoming, we propose a model based on LSTM and CNN which has cross-layer connections between the LSTM layers and the convolution layers. This model can simultaneously learn latent sequential information and capture important local features from time series features. Moreover, we introduce a method of constructing new features by training an XGBoost model on the existing features. Experimental results on the real-life dataset show that the model we proposed performs better than other comparison models.

CAPTCHA recognition based on deep convolutional neural network.

Aiming at the problems of low efficiency and poor accuracy of traditional CAPTCHA recognition methods, we have proposed a more efficient way based on deep convolutional neural network (CNN). The Dense Convolutional Network (DenseNet) has shown excellent classification performance which adopts cross-layer connection. Not only it effectively alleviates the vanishing-gradient problem, but also dramatically reduce the number of parameters. However, it also has caused great memory consumption. So we improve and construct a new DenseNet for CAPTCHA recognition (DFCR). Firstly, we reduce the number of convolutional blocks and build corresponding classifiers for different types of CAPTCHA images. Secondly, we input the CAPTCHA images of TFrecords format into the DFCR for model training. Finally, we test the Chinese or English CAPTCHAs experimentally with different numbers of characters. Experiments show that the new network not only keeps the primary performance advantages of the DenseNets but also effectively reduces the memory consumption. Furthermore, the recognition accuracy of CAPTCHA with the background noise and character adhesion is above 99.9%.

Qubit Neural Tree Network With Applications in Nonlinear System Modeling

This paper proposed a newly quantum-inspired Qubit neural tree network with improved Qubit neuron, cross-layer connections and distinct phase operation functions for each neurons. A hybrid evolutionary algorithm that combines the modified gene expression programming with particle swarm optimization is also introduced to obtain the optimal structure with related parameters of the Qubit neural tree network. Three nonlinear system modeling problems are selected to evaluate the effectiveness and performance of the proposed model. The simulation results indicate that the Qubit neural tree network has better nonlinear mapping and generalization ability than related methods do.