At present, automatic image matting methods primarily focus on portraits and hard edge targets, which gives them a limited ability to deal with low-resolution, complex, and blurred non-portrait targets. To address these issues, the current paper proposes an automatic image matting method called TGMatting. This method automatically optimizes Trimap generation through three modules: the U²S-Net pre-segmentation module, which is based on the U²-Net network, enhances segmentation by removing null convolutions and reducing oversampling interference; the BGTrimap module, which is also based on U²-Net, refines edge regions using optimized dilation-erosion methods and Manhattan distance for seed point sparsification, thus ensuring accurate region growth and background information removal; and in the last module, edges are transformed into mixed-pixel regions using Sobel operator and non-local-means denoising binarization, and a Trimap map is automatically generated by combining OTSU segmentation with pre-segmentation results, thus achieving fully automated processing. Finally, a transparency mask is obtained via FBAMatting, which enables interaction-<br/>free automatic matting. The experimental results demonstrate that the improved U²S-Net network reduces MAE by 0.003 on the SOD-Spider test set, enhances accurate detection of significant regions in low-resolution images compared to U²S-Net, and reduces BGTrimap&quots sum of absolute difference value by about 10% compared to other Trimap generation methods in IFMatting and KNN Matting.

To this point, there has been extensive research investigating human-robot motion retargeting, but the vast majority of existing methods rely on sensors or multiple cameras to detect human poses and movements, while many other methods are not suitable for usage on real-time scenarios. The current paper presents an integrated solution for performing realtime human-to-robot pose retargeting utilizing only regular monocular images and video as input data. We use deep learning models to perform three-dimensional human pose estimation on the monocular images and video, after which we calculate a set of joint angles that the robot must utilize to reproduce the detected human pose as accurately as possible. We evaluate our solution on Softbank’s NAO robot and show that it is possible to reproduce promising approximations and imitations of human motions and poses on the NAO robot, although it is subject to the limitations imposed by the robot’s degrees of freedom, joint constraints, and movement speed limitations.

Flip orthogonal frequency division multiplexing (OFDM) is a variation of OFDM which modifies the bipolar OFDM signal in the unipolar signal by flipping the negative sign of the subcarriers. Flip-OFDM in multiple input multiple output (MIMO) visible light communication (VLC) system improves the orthogonally of the subcarriers and reduces bit error rate, which results in a higher data rate and a more robust communication system. Machine learning (ML) and deep learning (DL) techniques are being used to improve various aspects of VLC systems such as modulation, channel estimation, and MIMO design, which can result in more robust and efficient communication systems. In this paper, deep neural network (DNN), convolution neural network (CNN) and long short-term memory (LSTM) algorithms are used to analyze flip-OFDM optical MIMO VLC system. The MIMO techniques, repetitive coding (RC), spatial modulation (SM), generalized spatial modulation (generalized-SM) and spatial multiplexing (SMP) are analyzed with and without flip-OFDM. Simulation results showed that generalized-SM outperformed SM, SMP and RC with and without flip-OFDM. In both scenarios, CNN improved performance and outperformed LSTM and DNN.

Text visualization is a complex technique that helps in data understanding and insight, and may lead to loss of information. Through the proposed low-dimensional vector representation learning method, deep learning and visualization through low-dimensional vector space construction were simultaneously performed. This method can transform a task-oriented dialogue dataset into low-dimensional coordinates, and based on this, a deep learning vector space can be constructed. The low-dimensional vector representation deep learning model found the intent of a sentence within a dataset and predicted the sentence components well in 3 out of 5 datasets. In addition, by checking the prediction results in the low-dimensional vector space, it was possible to improve the understanding of the data, such as identifying the structure or errors in the data.

The objective of this study was to describe a noisy threshold group testing model where positive and negative cases could occur depending on virus concentration in coronavirus disease 2019 (COVID-19) diagnosis with output results flipped due to measurement noise. We investigated lower bounds for successful reconstruction of a small set of defective samples in the noisy threshold group testing framework. To this end, using Fano’s inequality, we derived the minimum number of tests required to find unknown signals with defective samples. Our results showed that the minimum number of tests on probability of error for reconstruction of unknown signals was a function of the defective rate and noise probability. We obtained lower bounds for performance of the noisy threshold group testing framework with respect to noise intervals. In addition, the relationship between defective rate of signals and sparsity of group matrices to design optimal noisy threshold group testing systems is presented.

Among the different types of skin cancer, melanoma is the one with the highest death rate. Therefore, the early detection of melanoma and the development of technologies that can assist in this task have become significantly important. Convolutional neural networks are one of the most popular skin cancer classification methods. However, most of the available skin cancer datasets include images with lesions that are hard to differentiate from healthy skin or with a high presence of hair that can occlude the lesion. This characteristics of the images makes it harder to extract lesion features. Therefore, utilizing segmentation to extract the lesion location is an important step to reduce hair noise and improve lesion analysis. In this paper, two combining methods for segmentation and classification were explored: concatenation and multiplication. By utilizing these methods, it was possible to improve the accuracy of different neural network architectures by around 1% when compared to unmodified models without segmentation. The best-performing model was selected for further training. This model in conjunction with the segmentation module allowed for the correct re-classification of around 10% of the total examples in the dataset, indicating that a segmentation phase leads to an overall accuracy improvement and suggested that by improving the segmentation, an improvement on the overall accuracy can be obtained.

Deoxyribonucleic acid (DNA) is the basic entity that carries genetic instructions. This information is used in the evolution, progression, and improvement of all species. It is estimated that 10 CD-ROMs are required to store the genomic data of an individual being. With the increase in DNA sequencing equipment, an extensive heap of genomic data is created. The increase in DNA data in public databases is surpassing the rate of growth in storage space, thereby raising a significant concern related to data storage, transmission, retrieval, and search. To reduce the data storage and storage expense, lossless compression procedures were applied. Conventional compression methods are not proficient while compressing the biological data. Hence, several unique and contemporary lossless compression mechanisms were used to achieve improved compression ratio in biological sequences. Here, we scrutinize the diverse existing compression procedures that are appropriate for the compression of DNA sequences. The efficiency of algorithms is compared in terms of compression ratio, the ratio of the capacity of the compressed folder, and compression/decompression time. Main challenges and future research directions in DNA compression are also presented. Emphasis has been given to special references related to contemporary techniques.

Deep learning has become one of the most powerful prediction approaches, and it can be used to solve classification and regression problems. We present a novel deep learning-based indoor Wi-Fi path loss modeling approach. Specifically, we propose a local area multi-line scanning algorithm that generates input images based on measurement locations and a floor plan. As the input images contain information regarding the propagation environment between the fixed access points (APs) and measurement locations, a convolutional neural network (CNN) model can be trained to learn the features of the indoor environment and approximate the underlying functions of the Wi-Fi signal propagation. The proposed deep learning-based indoor path loss model can achieve superior performance over 3D ray-tracing methods. The average root mean square error (RMSE) between the predicted and measured received signal strength values in the two scenarios is 4.63 dB.

In this study, we have developed a point cloud segmentation algorithm for a collision avoidance system between cranes and other objects in construction yards. We used the Dynamic Graph CNN (DGCNN) algorithm to segment the point cloud of the entire yard into crane parts and backgrounds. The point cloud data were obtained from several LIDAR sensors attached to the crane. All points were grouped into specific core clusters using the DBSCAN algorithm. The core clusters were used to train the DGCNN after labeling with corresponding part names. This network classified the point cloud into crane types and their part names. Experimental results show that the crane part segmentation performance of the suggested algorithm is accurate enough to be used for collision avoidance system. It is possible to estimate the pose of a crane by comparing the segmented point clouds with those of the CAD model.