Electroencephalogram (EEG) signal processing is essential for achieving accurate and efficient real-time edge computing, particularly in resource-constrained smart wearables. They demand lightweight and optimized solutions for rapid analysis and decision-making. However, the higher dimensionality of EEG signals introduces challenges of latency and computational cost on edge devices. This article presents E-OptEEG, a hybrid ensemble metaheuristic framework designed to bring edge-optimized EEG processing and feature selection specifically for low-powered devices. The E-OptEEG framework integrates the strengths of multiple evolutionary feature selection algorithms, leveraging an ensemble threshold voting scheme to combine their outputs and identify the most relevant features. Further, E-OptEEG employs a fruit-fly optimization algorithm with deep metric feature transformation based on cosine similarity to transform the refined feature set to a minimal latent space. E-OptEEG demonstrates its ability to identify and select the most relevant features, enabling effective emotion and sentiment analysis from EEG signals captured through edge-powered wearable devices. Experimental evaluations against state-of-the-art feature selection techniques on three different EEG-based behavior modeling datasets highlight the framework’s effectiveness, achieving an average accuracy exceeding 98% with an average accuracy improvement of 2.55%. The E-OptEEG framework exemplifies the potential for lightweight artificial intelligence solutions to enable real-time, resource-efficient decision-making in wearable electronics.

Chest X-ray is one of the most common tests in radiology and plays a vital role in helping physicians spot different chest conditions. This paper proposes a new model called CAT-UNet to segment lung masses in chest X-ray images. The CAT-UNet uses TransUNet as the leading architecture and mixes CNN and transformer as an encoder. The CNN uses ResNet50 as the backbone, embedding the coordinate attention (CA) block and four skip connections of different scales to improve the accuracy of finding shallow features. Vision Transformer (ViT), which applied the transformer structure, was used in our method to enhance the feature representation ability of images, and Atrous Spatial Pyramid Pooling (ASPP) was used to adjust the filter’s field-of-view and control the resolution of features. In the decoder, the Convolutional Block Attention Modules (CBAM) are embedded for upsampling so that the segmentation details can be better optimized. To evaluate the performance and generalizability of the proposed method, we conducted a 3-fold cross-validation experiment using 1914 chest X-ray images labeled by radiologists collected from the Department of Radiology at Dalin Tzu Chi Hospital, Taiwan. Experimental results show that the proposed CAT-UNet achieves 89.06% on Dice, 91.62% on sensitivity, 98.64% on specificity, and 95.15% on accuracy, outperforming U-Net, TransUNet, and Swin-UNet encoders.

The Electrocardiogram (ECG) test detects and records cardiac-related electrical activity of the heart. The ECG test identifies and documents cardiac-related electrical activity in the heart. The use of ECG signals for cardiovascular disease nursing as a crucial component of preoperative evaluation is increasing. ECG signals need to denoise and display in a clear waveform due to the numerous noises. We have introduced Compact Shortcut Denoising Auto-encoder (CS-DAE) neural network, which reduces the noise from ECG signals. The Compact Shortcut approach compresses the features passed through the shortcut layers, which lowers the operation’s memory needs and improves the noise reduction impact. In addition, the encoder and decoder process the Pixel-Unshuffled and Pixel-Shuffled, which effectively mitigates the feature loss caused by down-sampling and up-sampling operations. As a result, the CS-DAE algorithm decreases the computation and required memory size while maintaining higher accuracy. We have used MITDB and NSTDB datasets for training and testing the proposed CS-DAE model, resulting in the average Percentage of Root Mean Square Difference (PRD) being 46.30% and the improvement of Signal-to-Noise Ratio (SNRimp) being 10.50. In addition, we have designed VLSI architect ure for the proposed CS-DAE neural network to accelerate low hardware cost and less computation. The TUL PYNQTM-Z2 development platform runs the Verilog code, which is used for VLSI architecture and has the lowest power consumption of 1.65W.

3D-HEVC (The 3D Extension of High Efficiency Video Coding) is the newest 3D video coding standard, which enriches multimedia applications with the video format of multi-view plus depth. For the depth map coding in 3D-HEVC, the advanced coding tools enhance the coding efficiency of the depth map and the quality of the synthesized view. However, the time consumption and complexity of 3D-HEVC also increase significantly. This paper utilizes the characteristics of human visual system to propose a fast algorithm based on visual perception for the acceleration of the depth intra coding of 3D-HEVC. The depth map is segmented into different regions by Otsu's auto-thresholding. The dominate edge direction is categorized for each prediction unit. We detect the perceptual edge based on just noticeable depth difference model to extract the area that may affect the visual perception. According to depth map segmentation and edge distribution, we reduce the corresponding intra angular modes and determine whether to perform depth modelling mode. We also incorporate the boundary continuity and rate-distortion cost thresholding to propose the fast coding unit decision. The experimental results show that the proposed algorithm eliminates 53.09% of the depth coding time with only 0.15% BD-BR on average. The coding performance of the proposed algorithm outperforms the previous works significantly.

Underwater image processing has been shown to exhibit significant potential for exploring underwater environments. It has been applied to a wide variety of fields, such as underwater terrain scanning and autonomous underwater vehicles (AUVs)-driven applications, such as image-based underwater object detection. However, underwater images often suffer from degeneration due to attenuation, color distortion, and noise from artificial lighting sources as well as the effects of possibly low-end optical imaging devices. Thus, object detection performance would be degraded accordingly. To tackle this problem, in this article, a lightweight deep underwater object detection network is proposed. The key is to present a deep model for jointly learning color conversion and object detection for underwater images. The image color conversion module aims at transforming color images to the corresponding grayscale images to solve the problem of underwater color absorption to enhance the object detection performance with lower computational complexity. The presented experimental results with our implementation on the Raspberry pi platform have justified the effectiveness of the proposed lightweight jointly learning model for underwater object detection compared with the state-of-the-art approaches.

Most of real-world datasets are of mixed type including both numeric and categorical attributes. Unlike numbers, operations on categorical values are limited, and the degree of similarity between distinct values cannot be measured directly. In order to properly analyze mixed-type data, dedicated methods to handle categorical values in the datasets are needed. The limitation of most existing methods is lack of appropriate numeric representations of categorical values. Consequently, some of analysis algorithms cannot be applied. In this paper, we address this deficiency by transforming categorical values to their numeric representation so as to facilitate various analyses of mixed-type data. In particular, the proposed transformation method preserves semantics of categorical values with respect to the other values in the dataset, resulting in better performance on data analyses including classification and clustering. The proposed method is verified and compared with other methods on extensive real-world datasets.

Beacon‐based indoor positioning is popular in recent years. In this work, the authors aim to enhance the positioning accuracy by proposing signal power ranking (SPR) and solving related optimisation‐based deployment problem of beacons using wireless communication and Bluetooth 4.0 Bluetooth low‐energy network technologies. The authors first adopt grid‐based field to be the proposed deployment field. Second, they convert the received signal strength indicator (RSSI) to several levels called SPR. Third, an optimisation‐based model for deployment problem of beacons in indoor positioning is proposed on the basis of the above two considerations. The proposed model is to minimise the number of beacons required under some fundamental conditions including full coverage and full discrimination, respectively. Finally, the algorithm of simulated annealing is applied to solve the linear programming problem in this model. By the optimal results, the user can obtain a vector table of RSSI for each location efficiently in the test field. On the other hand, the user in the test field can receive the beacon RSSI value at the same time. In order to determine the user's location, the received beacon RSSI value is compared with the values in the vector table.

With recent advances in embedded processing capability, vision-based real-time fire detection has been enabled in surveillance devices. This paper presents an image-based fire detection framework based on deep learning. The key is to learn a fire detector relying on tiny-YOLO (You Only Look Once) v3 deep model. With the advantage of lightweight architecture of tiny-YOLOv3 and training data augmentation by some parameter adjusting, our fire detection model can achieve better detection accuracy in real-time with lower complexity in the training stage. Experimental results have verified the effectiveness of the proposed framework.

The original self-organizing map (SOM) was proposed in the context of processing numeric data. In previous studies, an extended SOM incorporating data structure distance hierarchies has been proposed to facilitate handling of categorical values. The model could take into account the semantics embed ded in categorical values via distance hierarchies. In addition to manual construction by domain experts, an approach to learning distance hierarchies from datasets has been devised. However, the proposed approach in the previous study was based on supervised learning which demands presence of a class attribute in the dataset. In real-world applications, class attribute may not be available. Thus, the supervised approach can be inapplicable. In this article, we present several methods of unsupervised learning of distance hierarchies so that neither are class attribute nor domain experts required in measuring similarity degree between categorical values. We then integrate the learned distance hierarchies with the extended SOM to facilitate the application to datasets without a class attribute. We conduct experiments to verify feasibility and compare performance of the proposed unsupervised-learning methods.

Wireless sensor network in the industrial environment [industrial wireless sensor network (IWSN)] has data delivery time constraint. Due to the dynamic routing and transmission collision, the data delivery time is unpredictable. The authors utilised time division multiple access (TDMA) MAC to avoid data collision and to provide bounded transmission delay. Moreover, a linear programming model is proposed to construct the TDMA schedules, which is focused on spatial reuse and fixed routing in IWSN. The objective function of the model is to minimise the time slot usage to increase the overall network bandwidth. Finally, both simulated annealing algorithm and particle swarm optimisation are applied to approximate the optimal solution of time slot usage.

In recent years, ultrasound imaging was widely used in the diagnosis of lymph nodes. Most lymph nodes tend to have various internal echogenicities in the sonogram, which makes a definite diagnosis difficult. To overcome this problem, we propose a new image feature selection method based on ant colony optimization (ACO) for different imaging systems. The selected significant features are then applied to classify lymph node into six categories by support vector machine (SVM). Experimental results show that the proposed approach has high accuracy.

Ultrasound (US) imaging is more popular as a diagnostic tool than magnetic resonance imaging (MRI) and computerized tomography (CT) because it is inexpensive and easy to use. Most lymph nodes (LN) tend to have various internal echogenicities in the sonogram, which makes a definite diagnosis difficult. If the characteristic echogenicities for the major components of the lymph node can be identified, the interpretation of lymph sonography can be more accurate. In this paper, an ant colony optimization (ACO) algorithm is applied to select significant features from different ultrasound imaging systems for lymph node classification. The support vector machine (SVM) is employed to classify the lymph nodes into six categories. Experimental results show that the proposed approach achieve higher performance than those of other methods.

With level-set formulation, new contours can emerge during the evolution of contours. A defect inspection system that utilizes the evolution of zero-level contours for segmenting postsawing wafer is proposed in this study. The system utilizes a regional formulation, which improves the level-set segmentation in images with intensity inhomogeneity. An automatic threshold is used to set the initial contour to a contour near the die region. Fewer iterations are thus required to evolve the zero-level set to segment the wafer. Without the needs for filtering in advance, the inspection can be performed directly on the segmented results. The proposed approach outperforms other postsawing inspection methods in terms of accuracy.

Facial expression recognition is one of the most popular topics in emotion analysis. Most facial expression recognition systems are implemented using general expression models. Since facial expressions may be expressed differently by different people, inaccurate results are unavoidable. The proposed facial expression recognition system recognizes facial expressions using the facial features of an individual user. A radial basis function neural network is applied to classify seven emotions: neutral, happy, angry, surprised, sad, scared, and disgusted. Experiment results show that the proposed system can accurately identify emotions from facial expressions.

Facial expression are widely used for emotion recognition. Facial expressions may be expressed differently by different people subjectively, inaccurate results are unavoidable. Nevertheless, physiological reactions are non-autonomic nerves in physiology. The physiological reactions and the corresponding signals are hardly to control while emotions are excited. Therefore, an emotion recognition system with consideration of physiological signals is proposed in this paper. A specific designed mood induction experiment is performed to collect physiological signals of subjects. Five biosensors including electrocardiogram, respiration, galvanic skin responses (GSR), blood volume pulse, and pulse are used. Then a Support Vector Regression (SVR) is used to train three regression curves of three emotions (sad, fear, and pleasure). Experimental results show that the proposed method based on SVR emotion recognition has a good performance in accuracy.

Gender recognition is a hot research topic in recent years. Human-machine interfaces or video surveillance can be greatly improved if human gender can be recognized automatically. In this study, an embedded hidden Markov model is used for gender recognition. Video, which is recorded in different angles of view, is utilized to sample properties of each gender. Ten consecutive gait frames are segmented and organized as a composite image, which is used to establish EHMM. For video in each angle of view, two EHMMs are built and trained. The gender of the subject of a testing composite image is decided by the EHMM whose likelihood is most similar to the testing EHMM. We test the proposed approach using the CASIA Gait Database (Dataset B) in this study. Experimental results show that the proposed system can identify the gender of human accurately.

With level set formulation, new contours can system that utilizes the evolution of zero-level contours for segmenting post-sawing wafer is proposed in this study. The system also utilizes a regional formulation, which improves the level set segmentation in images with intensity inhomogeneity. With a proper configuration, the initial contour can be given in arbitrary positions in the background. Hence, a small fixed initial contour is applied to evolve the zero-level set to automatically segment the wafer. Without necessity of applying any filtering in advance, the inspection can be performed directly on the segmented results. The experimental results demonstrate the effectiveness of the proposed approach.

An emotion recognition system with consideration of facial expression and physiological signals is proposed in this paper. A specific designed mood induction experiment is performed to collect facial expressing images and physiological signals of subjects. We detected 14 feature points and extracted 12 facial features from facial expression images. Meanwhile, we measure the skin conductivity, finger temperature and heart rate from the subject. Both facial and physiological features are adopted to train the classifiers. Two learning vector quantization (LVQ) neural networks were applied to classify four emotions: love, joy, surprise and fear. Experimental results show the proposed recognition system is able to identify four emotions by facial expressions, physiological signals, and both of them.

Wafer defect inspection is an important process that is performed before die packaging. Conventional wafer inspections are usually performed using human visual judgment. A large number of people visually inspect wafers and hand-mark the defective regions. This requires considerable personnel resources and misjudgment may be introduced due to human fatigue. In order to overcome these shortcomings, this study develops an automatic inspection system that can recognize defective LED dies. An artificial neural network is adopted in the inspection. Actual data obtained from a semiconductor manufacturing company in Taiwan were used in the experiments. The results show that the proposed approach successfully identified the defective dies on LED wafers. Personnel costs and misjudgment due to human fatigue can be reduced using the proposed approach.

A system for the automatic inspection of LED wafer defects is proposed to detect defective dies in a four-element (aluminum gallium indium phosphide, AlGaInP) wafer. There are over 80000 dies on an LED wafer. Defective dies are typically visually identified with the aid of a scanning electron microscope. This process involves dozens of operators or engineers visually checking the wafers and hand marking the defective dies. However, wafers may not be fully and thoughtfully checked, and different observers usually find different results. These shortcomings lead to significant labor and production costs. Therefore, a solution that consists of two Hopfield neural networks, of which one is used to identify the LED die regions and the other is used to cluster the die into three groups, is proposed to facilitate the detection of defective dies in wafer images. The experimental results show that the proposed method successfully detects defective dies in a four-element wafer.

Most of the thyroid nodules are heterogeneous with various internal components, which confuse many radiologists and physicians with their various echo patterns in thyroid nodules. A lot of texture extraction methods were used to characterize the thyroid nodules. Accordingly, the thyroid nodules could be classified by the corresponding textural features. In this paper, five support vector machines (SVM) were adopted to select the significant textural features and to classify the nodular lesions of thyroid. Experimental results showed the proposed method classifies the thyroid nodules correctly and efficiently. The comparison results demonstrated that the capability of feature selection of the proposed method was similar to the sequential floating forward selection (SFFS) method. However, the proposed method is faster than the SFFS method.

The objective of this paper is to provide a complete solution to estimate the volume of the thyroid gland directly from US images. In this paper, the radial basis function (RBF) neural network is used to classify blocks of the thyroid gland; the integral region is further acquired by applying a specific region growing method to potential points. The parameters for evaluating the thyroid volume is estimated by a particle swarm optimization (PSO) algorithm. Experimental results of the thyroid region segmentation and volume estimation in US images show high potential of our proposed approach.

Digital watermarking is an important technique for protecting the intellectual property rights (IPR) of digital media. In this paper, we propose a DWT-based robust watermarking scheme with fuzzy adaptive resonance theory (Fuzzy ART) for still images. Experimental results demonstrate that the proposed scheme is robust against common image processing, geometric distortions and some intentional attacks. Without modifying the original image, the proposed scheme is not only a robust method but also a lossless one. Moreover, it is not necessary to use the original image during extracting the embedded watermark.

Recently, the watermarking is an important technique to protect copyright, which allows authentic watermark to be hidden in multimedia such as digital image, video and audio. In this paper, we propose a DWT-based counter-propagation neural network (CPN) for digital audio watermarking. The db4 filter of the Daubechies wavelet is applied in this paper. The coefficients obtained from 4-level db4 and the corresponding watermark is used for training the CPN. Different from the traditional methods, the watermark is embedded in the synapses of CPN instead of the original audio signal. In addition, because of the watermark is memorized in the synapses, most of the attacks are not degrade the quality of the extracted watermark image. Moreover, the watermark embedding procedure and extracting procedure are integrated into the proposed CPN. Experimental results show that the proposed method has capabilities of robustness, inaudibility and authenticity.

The objective of this paper is to develop a complete solution for recurrent nasal papilloma (RNP) detection. Recently, the gadolinium-enhanced dynamic magnetic resonance image (MRI) has been developed and widely used in clinical diagnosis of recurrent nasal papilloma. Owing to the response of RNP regions in gadolinium-enhanced magnetic resonance images is different from the response of normal tissues, the difference between the dynamic-MR images before and after administering contrast material can be used to extract the coarse RNP regions automatically. Then, a fuzzy algorithm for learning vector quantization (FALVQ) neural network is used to pick the suspicious RNP regions. Finally, a feature-based region growing method is applied to recover the complete RNP regions. The experimental results show that the proposed method can detect RNP regions automatically, correctly and fast.

Semiconductor wafer defect inspection is an important process before die packaging. The defective regions are usually identified through visual judgment with the aid of a scanning electron microscope. Dozens of people visually check wafers and hand-mark their defective regions. By this means, potential misjudgment may be introduced due to human fatigue. In addition, the process can incur significant personnel costs. Prior work has proposed automated post-sawing wafer defect inspection that is based on supervised neural networks. Since it requires learned patterns specific to each application, its disadvantage is the lack of product flexibility. Self-Organizing Neural Networks (SONNs) have been proven to have the capabilities of unsupervised auto-clustering. In this paper, automated wafer inspection based on a self-organizing neural network is proposed. Based on real-world data, experimental results show that the proposed method successfully identifies the defective regions on wafers with good performances.

Digital watermarking is an important technique for protection and identification that allows authentic watermarks to be hidden in multimedia such as image, audio, and video. Watermarking has been developed to protect digital media from being illegally reproduced and modified. Embedding and extracting watermark used to require complex procedures. In this paper, we propose a novel method called full counter-propagation neural network (FCNN) for digital image watermarking, in which the watermark is embedded and extracted through specific FCNN. Different from the traditional methods, the multiple cover images and the watermark are embedded in the synapses of a FCNN simultaneously instead of the cover images. Therefore, the watermarked image is almost the same as the original cover image. In addition, most of the attacks could not degrade the quality of the extracted watermark image. The experimental results show that the proposed method is able to achieve robustness, imperceptibility and authenticity in watermarking.

The special design of a Hopfield neural network, called contextual Hopfield neural network (CHNN), is presented for finding the edges of CT and MRI images. Different from conventional 2D Hopfield neural networks, the CHNN maps the 2D Hopfield network at the original image plane. With this direct mapping, the network is capable of incorporating pixel contextual information into a pixel's labeling procedure. As a result, the effect of tiny details or noises will be effectively removed by the CHNN and the drawback of disconnected fractions can be overcome. Furthermore, the problem of satisfying strong constraints can be alleviated and results in a fast converge. Our experimental results show that the CHNN can obtain more appropriate, more continued edge points than Laplacian-based, Marr-Hildreth's, Canny's, and wavelet-based methods.

Compared to object-based registration, feature-based registration is much less complex. However, in order for feature-based registration to work, the two image stacks under consideration must have the same acquisition tilt angle and the same anatomical location - two requirements that are not always fulfilled. In this paper, we propose a technique that reconstructs two sets of medical images acquired with different acquisition angles and anatomical cross sections into one set of images of identical scanning orientation and positions. The space correlation information among the two image stacks is first extracted and is used to correct the tilt angle and anatomical position differences in the image stacks. Satisfactory reconstruction results were presented to prove our points.

This paper presents an application of neural networks for solving the job-shop scheduling problem with the consideration of lot sizes (i.e. job batch sizes), which are important since jobs are often processed in batches. The energy-based neural network that have been proposed to solve this problem usually take a long time to converge to solutions. The authors previously (1994) proposed a new neural model which needs no special convergence procedure and can find optimal or near-optimal solutions of the problem at a much faster speed. However, in this model as well as other energy-based models, the number of neurons are proportional to the lot sizes of the jobs. This may complicate the implementation. In this paper, we extend our model to solve this problem. In this extended model, the number of neurons are fixed for different lot sizes. These results are quite good in terms of quality and speed. Furthermore, in this new model, mn(n+7) number of neurons are needed to solve an n-job m-machine problem with an arbitrary lot size for each job.