Trajectory-based writing system refers to writing a linguistic character or word in free space by moving a finger, marker, or handheld device. It is widely applicable where traditional pen-up and pen-down writing systems are troublesome. Due to the simple writing style, it has a great advantage over the gesture-based system. However, it is a challenging task because of the non-uniform characters and different writing styles. In this research, we developed an air-writing recognition system using three-dimensional (3D) trajectories collected by a depth camera that tracks the fingertip. For better feature selection, the nearest neighbor and root point translation was used to normalize the trajectory. We employed the long short-term memory (LSTM) and a convolutional neural network (CNN) as a recognizer. The model was tested and verified by the self-collected dataset. To evaluate the robustness of our model, we also employed the 6D motion gesture (6DMG) alphanumeric character dataset and achieved 99.32% accuracy which is the highest to date. Hence, it verifies that the proposed model is invariant for digits and characters. Moreover, we publish a dataset containing 21,000 digits; which solves the lack of dataset in the current research.

When sparsely sampled data are used to accelerate magnetic resonance imaging (MRI), conventional reconstruction approaches produce significant artifacts that obscure the content of the image. To remove aliasing artifacts, we propose an advanced convolutional neural network (CNN) called fully dense attention CNN (FDA-CNN). We updated the Unet model with the fully dense connectivity and attention mechanism for MRI reconstruction. The main benefit of FDA-CNN is that an attention gate in each decoder layer increases the learning process by focusing on the relevant image features and provides a better generalization of the network by reducing irrelevant activations. Moreover, densely interconnected convolutional layers reuse the feature maps and prevent the vanishing gradient problem. Additionally, we also implement a new, proficient under-sampling pattern in the phase direction that takes low and high frequencies from the k-space both randomly and non-randomly. The performance of FDA-CNN was evaluated quantitatively and qualitatively with three different sub-sampling masks and datasets. Compared with five current deep learning-based and two compressed sensing MRI reconstruction techniques, the proposed method performed better as it reconstructed smoother and brighter images. Furthermore, FDA-CNN improved the mean PSNR by 2 dB, SSIM by 0.35, and VIFP by 0.37 compared with Unet for the acceleration factor of 5.

Cancer remains a deadly disease. We developed a lightweight, accurate, general-purpose deep learning algorithm for skin cancer classification. Squeeze-MNet combines a Squeeze algorithm for digital hair removal during preprocessing and a MobileNet deep learning model with predefined weights. The Squeeze algorithm extracts important image features from the image, and the black-hat filter operation removes noise. The MobileNet model (with a dense neural network) was developed using the International Skin Imaging Collaboration (ISIC) dataset to fine-tune the model. The proposed model is lightweight; the prototype was tested on a Raspberry Pi 4 Internet of Things device with a Neo pixel 8-bit LED ring; a medical doctor validated the device. The average precision (AP) for benign and malignant diagnoses was 99.76% and 98.02%, respectively. Using our approach, the required dataset size decreased by 66%. The hair removal algorithm increased the accuracy of skin cancer detection to 99.36% with the ISIC dataset. The area under the receiver operating curve was 98.9%.

Air-writing is a growing research topic in the field of gesture-based writing systems. This research proposes a unified, lightweight, and general-purpose deep learning algorithm for a trajectory-based air-writing recognition network (TARNet). We combine a convolutional neural network (CNN) with a long short-term memory (LSTM) network. The architecture and applications of CNN and LSTM networks differ. LSTM is good for time series prediction yet time-consuming; on the other hand, CNN is superior in feature generation but comparatively faster. In this network, the CNN and LSTM serve as a feature generator and a recognizer, optimizing the time and accuracy, respectively. The TARNet utilizes 1-dimensional separable convolution in the first part to obtain local contextual features from low-level data (trajectories). The second part employs the recurrent algorithm to acquire the dependency of high-level output. Four publicly available air-writing digit (RealSense trajectory digit), character (RealSense trajectory character), smart-band, and Abas datasets were employed to verify the accuracy. Both the normalized and nonnormalized conditions were considered. The use of normalized data required longer training times but provided better accuracy. The test time was the same as those for nonnormalized data. The accuracy for RTD, RTC, smart-band, and Abas datasets were 99.63%, 98.74%, 95.62%, and 99.92%, respectively.

We propose a light-field microscopy display system that provides improved image quality and realistic three-dimensional (3D) measurement information. Our approach acquires both high-resolution two-dimensional (2D) and light-field images of the specimen sequentially. We put forward a matting Laplacian-based depth estimation algorithm to obtain nearly realistic 3D surface data, allowing the calculation of depth data, which is relatively close to the actual surface, and measurement information from the light-field images of specimens. High-reliability area data of the focus measure map and spatial affinity information of the matting Laplacian are used to estimate nearly realistic depths. This process represents a reference value for the light-field microscopy depth range that was not previously available. A 3D model is regenerated by combining the depth data and the high-resolution 2D image. The element image array is rendered through a simplified direction-reversal calculation method, which depends on user interaction from the 3D model and is displayed on the 3D display device. We confirm that the proposed system increases the accuracy of depth estimation and measurement and improves the quality of visualization and 3D display images.