This work proposes an approach for dynamic facial expression recognition to recognize emotions in controlled environments, due to its computational efficiency. The CREMA-D and RAVDESS datasets are used, from which sequences of 100 frames per video are extracted. Preprocessing for geometric features is performed using Face Mesh and facial alignment, while for deep features, face detection, facial alignment, resizing, and center cropping are applied. Geometric features are computed from the internal angles between facial landmarks, and deep features are extracted using MobileNetV2, ShuffleNetV2, and EfficientNet-B0, followed by dimensionality reduction via NCA. Both representations are concatenated and used as input to an LSTM (for CREMA-D) and a BiLSTM (for RAVDESS). The proposed method achieves UAR/WAR scores of 63.68%/63.71% on CREMA-D and 79.50%/80.21% on RAVDESS, demonstrating that the proposed approach is efficient and competitive without relying on architectures with higher computational cost.

Video Anomaly Detection (VAD) is a well-established area of research with significant potential for enhancing video surveillance in urban public transportation. However, current VAD systems often propose powerful methodologies but overlook their use in extreme environments like public transportation, necessitating a balance between performance and computational efficiency. In this paper, we evaluate a key component in many VAD frameworks: feature extractors. We investigate five extractors: Inflated 3D ConvNets (I3D), 3D Convolutional Neural Networks (C3D), Unified Transformer (UniFormer) in Small (UniFormer-S) and Base (UniFormer-B) versions, and Temporal Shift Module (TSM). These are integrated into a VAD architecture employing Bidirectional Encoder Representations from Transformers (BERT) with Multiple Instance Learning (MIL), chosen for its modularity and clear separation between the feature extractor and anomaly detector module. UniFormer-S demonstrated a processing rate of 4.64 clips per second with a computational demand of 28.717 GFLOPs on edge devices like the Jetson Orin NX (8GB RAM, 20W power). On the UCF-Crime dataset, UniFormer-S with BERT + MIL achieves an AUC of 79.74%. These findings highlight the promise of UniFormer-S and the use of edge devices like the Jetson Orin NX in public transportation due to their balance of performance and efficiency.

Automatic voice gender recognition (VGR) offers several real-world applications, including recommender system, human-robot interaction, and forensic application. VGR systems become challenging when these operate under unconstrained environments. In this study, we evaluate the performance of VGR systems using different fine-tuned pretrained Convolutional Neural Networks (CNNs), in which the speech signals under unconstrained environments are introduced as input data. First, preprocessing is applied to the original speech signal, which consists of noise attenuation based on low-pass filter and silence part removal based on sound amplitude. Then, the time-frequency features, such as Spectrogram, Mel-Spectrogram and Mel Frequency Cepstral Coefficients (MFCC) are extracted, which are converted into RGB images and processed by CNN models. Our research utilizes the VoxCeleb dataset, which is the largest video-audio dataset recorded under unconstrained environments. The results obtained by several fine-tuned CNN models provide higher accuracy compared with the state-of-the-art techniques on this topic. The best accuracy achieved is 98.58% using fine-tuned MobileNet, which is higher than the best accuracy provided by previous works.

In the last decade, the development and accessibility of image editing software have significantly increased, with many open-access tools available online. In this context, image-splicing attacks have gained popularity for their ability to manipulate images effectively. Consequently, there is a need to develop technologies capable of detecting splicing attacks using artificial intelligence systems. This paper introduces a method based on neural networks that identifies unique image features for detecting and localizing splicing manipulations by comparing the extracted features. The proposed approach used a neural network model with two identical branches to extract and compare features from the original and tampered images. Subsequently, the detected manipulated regions are divided into smaller blocks, and eigenvalues associated with each region are computed and compared with those of the original image. This process enables more efficient splicing detection. Experimental results demonstrate the method’s effectiveness against various splicing attacks, achieving a precision and efficacy of 90%. This study offers an innovative system for detecting and locating splicing image forgeries supported by comprehensive experimental validation.

The development of efficient audio coding schemes allowing an increasing information security and reduction of the storage requirement, is problem that has attracted the researchers interest over the last several years. To this end several schemes have been proposed, that allows an efficient compression and encryption of digital information. In most cases the information is firstly compressed before the encryption process. Because in several situations, to achieve a real time communication process, it is desirable to compress and encrypt the audio signal now when it is captured, it would be desirable to implement coding schemes able to encrypt and compress sensitive information simultaneously. A suitable approach to achieve this goal, is to use a compressing sensing-based systems which allows a simultaneous compression and encryption of the signal to be transmitted. This paper presents an audio encoding scheme using compressive sensing techniques which firstly segments the signal to be encoded signal in segments of M frames, each one with N samples. These are then transformed into a set of M sparse frames using the Discrete Cosine Transform (DCT) and multiplied by a properly designed random matrix of size M × N. The resulting vectors are then concatenated to generate the compressed and encrypted matrix. It is then feed to a chaotic mixing scheme to further increase the security of proposed system. Evaluation results shows that the proposed system achieves and efficient and secure compression and encryption, while satisfying the extended Wyner secrecy criterion (EWS).

This paper proposes an innovative deep learning algorithm, denominated as Neuroevolution of Hybrid Neural Networks in a Robotic Agent (NRNH-AR). This algorithm stands out in computational efficiency, evolutionary stability, and classification speed, surpassing other existing methods such as FAM-HGNN, I3A, DQN, among others. The triptych design of the NRNH-AR optimizes resource usage and exhibits superior adaptability in dynamic environments, which is a critical factor in robotics. The integration of artificial vision techniques enhances the capability for rapid classification of observations. Throughout 300 epochs in a dynamic environment, the importance of a minimal database for the algorithm’s effective evolution was evidenced. Moreover, it was found that the accuracy of the Convolutional Neural Network (CNN) directly impacts the performance of the NRNH-AR. An additional study of 850 epochs in a static environment provided a deeper understanding of the evolution and adaptation of learning networks. The NRNH-AR is positioned as a viable and promising solution in the field of applied deep learning in robotics.

Drowsiness and driver distraction are considered the main causes of traffic accidents in the world. Considering this situation, this paper proposes two important modifications to our previously proposed driver drowsiness and distraction detector for real-time implementation on handheld mobile devices, such as smartphones. The first modification is due to a large variation in the capacity of mobile devices. To adapt the proposed system to a wide range of mobile devices, we present two automatic threshold calculations, which are used to differentiate driver drowsiness from normal blinking and dangerous driver distraction from normal short-term distraction. The second modification is related to the alarm during a continuous dangerous situation of the driver. We introduce a new algorithm to ensure the continuous activation of the alarm while the dangerous situation continues. These improvements perform as the general algorithm, since when it was implemented in mobile devices with low computational power, as well as in devices that do not have these limitations, the alarm activation times were not affected; On the other hand, it was possible to increase the accuracy originally given by the first system with respect to Ground Truth by almost 25% on average, resulting in alarm activations not being affected to a great extent by the natural errors that the convolutional neural networks (CNN) may cause, these improvements are shown and supported by the implementation in real time through video links provided in this work.

Automatic detection of the referable Diabetic Retinopathy (RDR) has become essential in diabetic patients, especially who live in the remote regions, to avoid a serious visual impairment. For this reason, different approaches have been developed with the aim of detect and segment the principal DR lesions for automatic diagnosis of the RDR. Exudate is one of the DR lesions and if these lesions appear in the macular region, a diabetic macular edema (DME) can be suspected and a detailed analysis by ophthalmologist is required. Then it is important to detect these lesions with their position related to the macular region to determine the danger level. This paper presents an automatic method to localize the exudates and optic disc (OD) using Single Shot Detector (SSD) scheme based on MobileNet-V1 as base network to determine if the risk of DME exits to indicate patients the necessity of consultation by ophthalmologist. The proposed system is evaluated using MESSIDOR Database, providing 89.15% accuracy, 88.17% sensitivity and 91.67% specificity.

Considering that an accurate detection of infected mosquitos may directly avoid the propagation of mosquito-borne disease; in this paper, we propose a detection system of infected mosquitos by Dengue virus type II, that uses seven spectral feature measures, which are applied to the spectrogram estimated from wingbeat signal emitted by mosquito’s flight. To evaluate the proposed system, we construct our own dataset with 20 infected Aedes aegypti by Dengue and 20 healthy ones. Seven spectral analysis methods, such as Spectral Rolloff, Spectral Centroide, etc., are applied to the spectrogram obtained by using the Short Time Fourier Transform (STFT) to generate feature vectors with 15 elements. These are feed into common machine learning techniques, such as Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and Logistic Regression to detect the infected mosquitos differentiating form the healthy ones. Evaluation results show that, the best detection accuracy (84.32%) is provided by the KNN with K=3.

In recent years, the reversible data hiding techniques also known as lossless or invertible data hiding, has gradually become a very active research area. The reversibility of these schemes makes possible to extract the embedded data without errors, as well as to restore the cover medium to its original state. Furthermore, to guarantee the security and confidentiality of the hidden data and the image, reversible data hiding schemes over encrypted domain are presented as a promising solution to solve several issues of information security. This paper presents a study case of reversible data hiding schemes over encrypted domain oriented to the protection of the sharing and distribution of color images hosted in cloud storage services. The experimental results are presented in terms of imperceptibility, capacity, confidentiality, and visual quality, respectively.

The development of acoustic scenes recognition systems has been a topic of extensive research due to its applications in several fields of science and engineering. This paper proposes an environmental system in which firstly a time-frequency representation is obtained using the Continuous Wavelet Transform (CWT). The time frequency representation is then represented as a color image using the Viridis color map, which is then inserted into a Deep Neural Network (DNN) to carry out the classification task. Evaluation results using several public data bases show that proposed scheme provides a classification performance better than the performance provided by other previously proposed schemes.

In this paper, we propose Long-Short Term Memory (LSTM)-based mosquito’s genus classification, in which the time-frequency features are extracted from the wingbeat sound of mosquitos of three genera, Aedes, Anopheles and Culex. The extracted features are fed into the proposed LSTM-based classifier. We evaluated three time-frequency features, which are: Mel Spectrogram, Log-Mel spectrogram, and Mel-frequency Cepstral Coefficients (MFCC). The proposed scheme is composed by two LSTM layers and one Fully Connected layer connected to a SoftMax activation function. The classification accuracies using the three features are 92.97(±0.2)%, 96.71(±0.2)% and 96.65(±0.2)%, respectively. The Area Under Curve (AUC) of the Receiver Operating Characteristics (ROC) for each feature are also obtained, which are 0.9944, 0.9986 and 0.9987, respectively. The proposed classifier requires approximately 62,000 trainable parameters. This number is much smaller than that required for the state-of-arts CNNs, such as AlexNet and Vgg16. This compact configuration of the proposed scheme takes advantage of the mobile and IoT implementation, because the number of trainable parameters is directly proportional to the amount of memory and CPU required.

In this paper, a classification of mosquito’s specie is performed using mosquito wingbeats samples obtained by optical sensor. Six world-wide representative species of mosquitos, which are Aedes aegypti, Aedes albopictus, Anopheles arabiensis, Anopheles gambiae and Culex pipiens, Culex quinquefasciatus, are considered for classification. A total of 60,000 samples are divided equally in each specie mentioned above. In total, 25 audio feature extraction algorithms are applied to extract 39 feature values per sample. Further, each audio feature is transformed to a color image, which shows audio features presenting by different pixel values. We used a fully connected neural networks for audio features and a convolutional neural network (CNN) for image dataset generated from audio features. The CNN-based classifier shows 90.75% accuracy, which outperforms the accuracy of 87.18% obtained by the first classifier using directly audio features.

This paper shows the feasibility of a variant of the Generative Adversarial Network (GAN), called Star GAN, for music genre transfer. This method is noteworthy in that it simultaneously learns many-to-many mappings across different attribute domains using a single generator network. A similar architecture to research in MuseGAN and CycleGAN is applied. Also, as in MGTGAN, Desert Camel MIDI dataset is use for training and testing.

In this paper, a novel Deep Neural Network topology is presented with the objective of recognizing the Aedes aegypti and Aedes albopictus mosquito in their larvarian stage, which are the vectors that cause Dengue, Chikungunya, Zika and Yellow Fever outbreaks. This solution allows to determine if a sample image is a larva of the Aedes aegypti or Aedes albopictus mosquito with an accuracy of 91.28%, a true positive rate of 94.18% and a true negative rate of 88.37%. This Deep Neural Network topology allows the implementation of fast and accurate preventive measures in under-developed countries and isolated areas where a trained specialist might not be available.

This paper proposes a novel methodology for change detection in video sequences, which consists in the use of projection of the first eigenvector over the current frame in the video sequence. These eigenvectors are computed using the Incremental Principal Component Analysis (IPCA), assuming that the incremental computation of the eigenvalues and eigenvectors is made using the incremental block approach considering only two frames i.e. the past and the current frames in each incremental block. The main contribution of this work, is the use of the idea that the first eigenvector projects the maximum variability in their data matrix and then by using the incremental block of two frames in the IPCA, the maximum variability in those images could be considered as the change between them; such that after the post-processing in the projected matrix, we are able to labeled the change between the past and the current frames.

This paper proposed a removable visible watermark system for video sequences using dual watermarking technique, embedding both visible and invisible watermarks into the video sequence. The visible watermark is embedded into every frame of the video sequence in the Discrete Cosine Transform (DCT) domain, considering the Human Visual System (HVS) model. The invisible watermark is embedded into every intra-frame during the MPEG encoding, using the Quantization Index Modulation-Dither Modulation (QIM-DM) technique. In the proposed scheme, two user's keys are introduced to ensure that only authorized users can remove the visible watermark to obtain clear video sequence. The experimental results show the desirable performance of the proposed scheme, in which unauthorized users cannot remove the visible watermark. The process is carried out through a totally blind manner without any extra information.

In this paper, fourteen descriptors are evaluated for urban-rural classification of aerial images. Among these fourteen descriptors, eleven descriptors consist of texture-based, color-based con combination of these two descriptors. Rest three descriptors are based on dictionaries generated using the Lempel-Ziv-Welch (LZW) data compression algorithm. The classification is carried out using Support Vector Machine (SVM) with radial basis function as kernel function and KNN algorithm. The performance of these images descriptors are evaluated using accuracy, precision, sensitivity and specificity. From evaluation results, we conclude the Gabor descriptor combined with Dominant Color descriptor provides better performance, obtaining its accuracy more than 91%.

The facial expression recognition has been a topic of active researches given a result the proposal of several efficient algorithms; however, in most cases they remain limited to controlled conditions situations. In this study, we tackle the challenge of recognizing emotions through the facial expression into activities in-the-wild adding the accuracy rate for each expression. To this end we an algorithm that allows accurate face expression recognition in an uncontrolled environment, that means different kind of illumination, backgrounds, occlusions, face's profiles, etc. Proposed system firstly detects different profile of face (left, frontal and right), Then it uses only the frames in which the face profile is frontal, in the next step the face regions of interest (ROI) are segmented automatically to carry out the feature extraction. We use a classifier based on clustering, it has the advantage that if a new class (emotion) is added, it is not necessary to train this completely. Proposed system was evaluated using short video clips of several pictures together with description sentences describing the main activity in the video. The evaluation results show that the proposed scheme is able to recognize the face's profiles with the recognition rate to approximately 93% and principal emotions in unconstrained video sequences.

In this paper, we propose software protection algorithm against illegal copy by unauthorized persons using software watermarking. In the proposed scheme, the software license information, related to the media access control (MAC) address and software product-key, is embedded into the software logotype image as a watermark sequence. At the first stage of the software execution, the watermark sequence is extracted from the logotype and compared with the current MAC address of the hardware devise. If the extracted MAC address differs from the hardware's MAC address, then the software displays a warning message related to the invalid license and finalizes its execution, otherwise the software continues its normal operation. The performance of the proposed scheme is analyzed from the difficulty of software piracy and reverse engineering points of view, and the evaluation results show the effectiveness of the proposed algorithm.

In this paper we propose a visible watermarking algorithm, in which a visible watermark is embedded into the Bayer Colour Filter Array (CFA) domain. The Bayer CFA is the most common raw image representation for images captured by single sensor digital cameras equipped in almost all mobile devices. In proposed scheme, the captured image is watermarked before it is compressed and stored in the storage system. Then this method enforces the rightful ownership of the watermarked image, since there is no other version of the image rather than the watermarked one. We also take into consideration the Human Visual System (HVS) so that the proposed technique provides desired characteristics of a visible watermarking scheme, such that the embedded watermark is sufficiently perceptible and at same time not obtrusive in colour and grey-scale images. Unlike other Bayer CFA domain visible watermarking algorithms, in which only binary watermark pattern is supported, proposed watermarking algorithm allows grey-scale and colour images as watermark patterns. It is suitable for advertisement purpose, such as digital library and e-commerce, besides copyright protection.

Recently hierarchical threshold secret image sharing (SIS) scheme was proposed to provide hierarchical access structures to the conventional SIS. These hierarchical access structures are more adequate than the conventional SIS in many practical situations. However in the actual hierarchical SIS scheme, even if the access condition is not satisfied, the secret image is visually leaked, which is very serious security deficiency of this SIS scheme. The present work proposes a new condition for the hierarchical threshold SIS scheme to guarantee the secrecy of the secret image when the access condition is not satisfied. The experimental results show the functionality of the SIS with the proposed condition, in which if the access condition is satisfied then the secret image is recovered in lossless manner, otherwise only pseudorandom pattern is extracted.

Video transcoding is a legitimate operation widely used to modify video format in order to access the video content in the end-user's devices, which may have some limitations in the spatial and temporal resolutions, bit-rate and video coding standards. In many previous watermarking algorithms the embedded watermark is not able to survive video transcoding, because this operation is a combination of some aggressive attacks, especially when lower bit-rate coding is required in the target device. As a consequence of the transcoding operation, the embedded watermark may be lost. This paper proposes a robust video watermarking scheme against video transcoding performed on base-band domain. In order to obtain the watermark robustness against video transcoding, four criteria based on Human Visual System (HVS) are employed to embed a sufficiently robust watermark while preserving its imperceptibility. The quantization index modulation (QIM) algorithm is used to embed and detect the watermark in 2D-Discrete Cosine Transform (2D-DCT) domain. The watermark imperceptibility is evaluated by conventional peak signal to noise ratio (PSNR) and structural similarity index (SSIM), obtaining sufficiently good visual quality. Computer simulation results show the watermark robustness against video transcoding as well as common signal processing operations and intentional attacks for video sequences.

In this work we propose a parallel implementation of a Cellular Automata (CA)-based multiple secret image sharing (SIS) with a Least Significant Bit (LSB) Stenography. In previous CA-based multiple-SIS schemes, the resulting encrypted files are noise-like appearance images, which arises suspicion of observers when these are transmitted through a public network. To solve this problem, the LSB steganography is used to hide the encrypted noise-like images into camouflage images. To reduce execution time of the encoding and the decoding process of the SIS scheme, a parallel implementation based on CUDA technology is proposed. The simulation results demonstrate that the proposed parallel algorithm is more than 7 times faster than the conventional sequential algorithm. This reduction of temporal complexity and obtaining unnoticed shares allow the reliability of this scheme to be used in many information security fields such as the user-devices and cloud storage environments.

This paper propose a new interest point detector, denominated Atomic Harris-Affine detector, in which the Gaussian function is replaced by a 2D up(x,y) Atomic function (AF). Proposed scheme takes advantage of the desirable characteristics of AF, such as compact support and low spectral leakage. The simulation results show that proposed detector improves repeatability and convergence rate on, 15% and 20%, respectively, compared with the conventional one under several image conditions, such as illumination, viewpoints, blurring level, rotation angle, scaling factor and JPEG lossy compression rate. Also the detection accuracy of proposed detector allows a faster detection of interest points compared with the conventional one.

In this paper, a robust watermarking method against geometric distortions and common signal processing is proposed. This method is based on two modifications in the 2D histograms. In the first modification, a selected region of 2D histogram composed by R-G color components is modified according to the watermark pattern. In the second modification, another 2D histogram composed by B component and image feature is dynamically partitioned to embed the watermark pattern. The experimental results show robustness against several geometric distortions, common signal processing attacks and some combination attacks. Also the comparison with the prior methods shows the better performance of the proposed method.

A Real-Time audio watermarking scheme is presented. The strength of audio signal modifications is limited by the necessity to produce an output signal that is perceptually equal to the original signal. The proposed scheme uses a blind detection approach. To embed the watermark a spread spectrum algorithm is used in the Modulated Complex Lapped Transform (MCLT) domain. Here the watermark is generated using a private key and modeling according to the Human Auditory System. Comparison of several whitening methods in the receptor side is carried out. The embedded watermark is robust to common attacks such like, D/A A/D conversion, filtering, noise addition and high quality MPEG audio coding.

A Real-Time audio watermarking system is presented. System contains two circuits: the marker circuit and the detector circuit. Watermark detection is blind, so, there is no necessary original signal to detect embedded information. Watermark detection does not need a synchronization stage; therefore, detection is carried out on any time. Both, marker and detector are implemented in a commercial Digital Signal Processor Starter Kit from Texas Instruments and are fully host-independent.

This paper proposes a time varying normalized LMS (TVS-NLMS) algorithm for adapting an echo canceler structure. Proposed algorithm reduces distortion during double-talk, without increasing the computational cost nor decreasing the convergence rate of the normalized LMS algorithm significantly. Simulation results confirm the desirable features of the proposed schema.>