In this paper, we present a novel approach that aims to solve one of the main challenges in hand gesture recognition tasks in static images, to compensate for the accuracy lost when trained models are used to interpret completely unseen data. The model presented here consists of two main data-processing stages. A deep neural network (DNN) for performing handshape segmentation and classification is used in which multiple architectures and input image sizes were tested and compared to derive the best model in terms of accuracy and processing time. For the experiments presented in this work, the DNN models were trained with 24,000 images of 24 signs from the American Sign Language alphabet and fine-tuned with 5200 images of 26 generated signs. The system was real-time tested with a community of 10 persons, yielding a mean average precision and processing rate of 81.74% and 61.35 frames-per-second, respectively. As a second data-processing stage, a bidirectional long short-term memory neural network was implemented and analyzed for adding spelling correction capability to our system, which scored a training accuracy of 98.07% with a dictionary of 370 words, thus, increasing the robustness in completely unseen data, as shown in our experiments.

This paper reports the design and analysis of an American Sign Language (ASL) alphabet translation system implemented in hardware using a Field-Programmable Gate Array. The system process consists of three stages, the first being the communication with the neuromorphic camera (also called Dynamic Vision Sensor, DVS) sensor using the Universal Serial Bus protocol. The feature extraction of the events generated by the DVS is the second part of the process, consisting of a presentation of the digital image processing algorithms developed in software, which aim to reduce redundant information and prepare the data for the third stage. The last stage of the system process is the classification of the ASL alphabet, achieved with a single artificial neural network implemented in digital hardware for higher speed. The overall result is the development of a classification system using the ASL signs contour, fully implemented in a reconfigurable device. The experimental results consist of a comparative analysis of the recognition rate among the alphabet signs using the neuromorphic camera in order to prove the proper operation of the digital image processing algorithms. In the experiments performed with 720 samples of 24 signs, a recognition accuracy of 79.58% was obtained.

Diabetic retinopathy (DR) is the primary cause of blindness in developing and developed countries. Early-stage DR detection reduces the risk of blindness in Diabetes Mellitus (DM) patients. There has been a sharp rise in the prevalence of DM in recent years, especially in low- and middle-income countries. In this context, automated artificial intelligence-based DM screening is a crucial tool to help classify the considerable amount of Retinal Fundus Images (RFI). However, retinal image quality assessment has shown to be fundamental in real-world DR screening processes to avoid out-of-distribution data, drift, and images lacking relevant anatomical information. This work analyzes the spatial domain features and image quality assessment metrics for carrying out Deep Learning (DL) classification and detecting notable features in RFI. In addition, a novel lightweight convolutional neural network is proposed specifically for binary classification at a low computational cost. The training results are comparable to state-of-the-art neural networks, which are widely used in DL applications. The implemented architecture achieves 98.6% area under the curve, and 97.66%, and 98.33% sensitivity and specificity, respectively. Moreover, the object detection model trained achieves 94.5% mean average precision. Furthermore, the proposed approach can be integrated into any automated RFI analysis system.

This paper presents a platform that automatically generates custom hardware accelerators for convolutional neural networks (CNNs) implemented in field-programmable gate array (FPGA) devices. It includes a user interface for configuring and managing these accelerators. The herein-presented platform can perform all the processes necessary to design and test CNN accelerators from the CNN architecture description at both layer and internal parameter levels, training the desired architecture with any dataset and generating the configuration files required by the platform. With these files, it can synthesize the register-transfer level (RTL) and program the customized CNN accelerator into the FPGA device for testing, making it possible to generate custom CNN accelerators quickly and easily. All processes save the CNN architecture description are fully automatized and carried out by the platform, which manages third-party software to train the CNN and synthesize and program the generated RTL. The platform has been tested with the implementation of some of the CNN architectures found in the state-of-the-art for freely available datasets such as MNIST, CIFAR-10, and STL-10.

The synthetic aperture radar (SAR) tomography (TomoSAR) inverse problem is commonly tackled in the context of the direction-of-arrival estimation theory. The latter allows achieving super-resolution, along with ambiguity levels reduction, thanks to the use of parametric focusing methods, as multiple signal classification (MUSIC), and statistical regularization techniques, like the maximum-likelihood inspired adaptive robust iterative approach (MARIA). Nevertheless, in order to correctly suit the considered signal model, MUSIC and most regu-larization approaches require an appropriate setting of the involved parameters. In both cases, the accuracy of the retrieved solutions depends on the right selection of the assigned values. Thus, with the aim of dealing with such an issue, this article addresses several parameter selection strategies, adapted specifically to the TomoSAR scenario. Parametric techniques as MUSIC solve the TomoSAR problem in a different manner as the regularization methods do, hence, each approach demands different methodologies for the proper estimation of their parameters. Consequently, we refer to the Kullback-Leibler information criterion for the model order selection of parametric techniques as MUSIC, whereas we rather explore the Morozov's discrepancy principle, the L-Curve, the Stein's unbiased risk estimate and the generalized cross-validation, to choose the regularization parameters. After the incorporation of these criteria to MUSIC and MARIA, respectively, their capabilities are first analyzed through simulations, and later on, utilizing real data acquired from an urban area.

This work presents a configurable architecture for an artificial neural network implemented with a Field Programmable Gate Array (FPGA) in a System on Chip (SoC) environment. This architecture can reproduce the transfer function of different Multilayer Feedforward Neural Network (MFNN) configurations. The functionality of this configurable architecture relies on a single perceptron, multiplexers, and memory blocks that allow routing, storing, and processing information. The extended Kalman filter is the training algorithm that obtains the optimal weight values for the MFNN. The presented architecture was developed using Verilog Hardware Description Language, which permits designing hardware with a fair number of logical resources, and facilitates the portability to different FPGAs models without compatibility problems. A SoC that mainly incorporates a microprocessor and a FPGA is proposed, where the microprocessor is used for configuring the the MFNN and to enable and disable some functional blocks in the FPGA. The hardware was tested with measurements from a GaN class F power amplifier, using a 2.1 GHz Long Term Evolution signal with 5 MHz of bandwidth. In particular, a special case of an MFNN with two layers, i.e., a real-valued nonlinear autoregressive with an exogenous input neural network, was considered. The results reveal that a normalized mean square error value of −32.82 dB in steady-state was achievable, with a 71.36% generalization using unknown samples.