Light Convolutional Neural Network Research Articles

Evaluation of monolithic crystal detector with dual-ended readout utilizing multiplexing method

Objective. Monolithic crystal detectors are increasingly being applied in positron emission tomography (PET) devices owing to their excellent depth-of-interaction (DOI) resolution capabilities and high detection efficiency. In this study, we constructed and evaluated a dual-ended readout monolithic crystal detector based on a multiplexing method. Approach. We employed two 12 × 12 silicon photomultiplier (SiPM) arrays for readout, and the signals from the 12 × 12 array were merged into 12 X and 12 Y channels using channel multiplexing. In 2D reconstruction, three methods based on the centre of gravity (COG) were compared, and the concept of thresholds was introduced. Furthermore, a light convolutional neural network (CNN) was employed for testing. To enhance depth localization resolution, we proposed a method by utilizing the mutual information from both ends of the SiPMs. The source width and collimation effect were simulated using GEANT4, and the intrinsic spatial resolution was separated from the measured values. Main results. At an operational voltage of 29 V for the SiPM, an energy resolution of approximately 12.5 % was achieved. By subtracting a 0.8 % threshold from the total energy in every channel, a 2D spatial resolution of approximately 0.90 mm full width at half maximum (FWHM) can be obtained. Furthermore, a higher level of resolution, approximately 0.80 mm FWHM, was achieved using a CNN, with some alleviation of edge effects. With the proposed DOI method, a significant 1.36 mm FWHM average DOI resolution can be achieved. Additionally, it was found that polishing and black coating on the crystal surface yielded smaller edge effects compared to a rough surface with a black coating. Significance. The introduction of a threshold in COG method and a dual-ended readout scheme can lead to excellent spatial resolution for monolithic crystal detectors, which can help to develop PET systems with both high sensitivity and high spatial resolution.

Unleashing the power of Convolutional Neural Networks in license plate recognition and beyond

This article explores the application of Convolutional Neural Networks (CNN) in the field of license plate recognition. It begins by introducing the architecture of CNN, which consists of three key layers: Convolutional Layers, Pooling Layers, and Fully Connected Layers. The article then references three relevant papers that demonstrate how CNNs are applied in license plate recognition. The first paper utilizes TensorFlow to construct a CNN model and integrates it with an STM32MP157 embedded chip for license plate recognition. The second paper presents a real-time car license plate detection and recognition method called Multi-Task Light CNN, emphasizing robustness. The third paper employs the ResNet+FPN feature extraction network of the Mask R-CNN model and annotates a license plate dataset. The article highlights the promising future of CNNs in various fields beyond license plate recognition, emphasizing their potential for further development and industrial applications. CNNs have proven to be versatile and powerful tools in computer vision, offering solutions to a wide range of problems. Their adaptability and effectiveness make them a key player in the ongoing advancement of artificial intelligence and automation technologies.

A Refined Wind Power Forecasting Method with High Temporal Resolution Based on Light Convolutional Neural Network Architecture

With a large proportion of wind farms connected to the power grid, it brings more pressure on the stable operation of power systems in shorter time scales. Efficient and accurate scheduling, operation control and decision making require high time resolution power forecasting algorithms with higher accuracy and real-time performance. In this paper, we innovatively propose a high temporal resolution wind power forecasting method based on a light convolutional architecture—DC_LCNN. The method starts from the source data and novelly designs the dual-channel data input mode to provide different combinations of feature data for the model, thus improving the upper limit of the learning ability of the whole model. The dual-channel convolutional neural network (CNN) structure extracts different spatial and temporal constraints of the input features. The light global maximum pooling method replaces the flat operation combined with the fully connected (FC) forecasting method in the traditional CNN, extracts the most significant features of the global method, and directly performs data downscaling at the same time, which significantly improves the forecasting accuracy and efficiency of the model. In this paper, the experiments are carried out on the 1 s resolution data of the actual wind field, and the single-step forecasting task with 1 s ahead of time and the multi-step forecasting task with 1~10 s ahead of time are executed, respectively. Comparing the experimental results with the classical deep learning models in the current field, the proposed model shows absolute accuracy advantages on both forecasting tasks. This also shows that the light architecture design based on simple deep learning models is also a good solution in performing high time resolution wind power forecasting tasks.

Complementary regional energy features for spoofed speech detection

Automatic speaker verification systems are found to be vulnerable to spoof attacks such as voice conversion, text-to-speech, and replayed speech. As the security of biometric systems is vital, many countermeasures have been developed for spoofed speech detection. To satisfy the recent developments on speech synthesis, publicly available datasets became more and more challenging (e.g., ASVspoof 2019 and 2021 datasets). A variety of replay attack configurations were also considered in those datasets, as they do not require expertise, hence easily performed. This work utilizes regional energy features, which are experimentally proven to be more effective than the traditional frame-based energy features. The proposed energy features are independent from the utterance length and are extracted over nonoverlapping time-frequency regions of the magnitude spectrum. Different configurations are considered in the experiments to verify the regional energy features’ contribution to the performance. First, light convolutional neural network – long short-term memory (LCNN – LSTM) model with linear frequency cepstral coefficients is used to determine the optimal number of regional energy features. Then, SE-Res2Net model with log power spectrogram features is used, which achieved comparable results to the state-of-the-art for ASVspoof 2019 logical access condition. Physical access condition from ASVspoof 2019 dataset, logical access and deep fake conditions from ASVspoof 2021 dataset are also used in the experiments. The regional energy features achieved improvements for all conditions with almost no additional computational or memory loads (less than 1% increase in the model size for SE-Res2Net). The main advantages of the regional energy features can be summarized as i) capturing nonspeech segments, ii) extracting band-limited information. Both aspects are found to be discriminative for spoofed speech detection.

ENAS-B: Combining ENAS With Bayesian Optimization for Automatic Design of Optimal CNN Architectures for Breast Lesion Classification From Ultrasound Images.

Efficient Neural Architecture Search (ENAS) is a recent development in searching for optimal cell structures for Convolutional Neural Network (CNN) design. It has been successfully used in various applications including ultrasound image classification for breast lesions. However, the existing ENAS approach only optimizes cell structures rather than the whole CNN architecture nor its trainable hyperparameters. This paper presents a novel framework for automatic design of CNN architectures by combining strengths of ENAS and Bayesian Optimization in two-folds. Firstly, we use ENAS to search for optimal normal and reduction cells. Secondly, with the optimal cells and a suitable hyperparameter search space, we adopt Bayesian Optimization to find the optimal depth of the network and optimal configuration of the trainable hyperparameters. To test the validity of the proposed framework, a dataset of 1522 breast lesion ultrasound images is used for the searching and modeling. We then evaluate the robustness of the proposed approach by testing the optimized CNN model on three external datasets consisting of 727 benign and 506 malignant lesion images. We further compare the CNN model with the default ENAS-based CNN model, and then with CNN models based on the state-of-the-art architectures. The results (error rate of no more than 20.6% on internal tests and 17.3% on average of external tests) show that the proposed framework generates robust and light CNN models.

Morse wavelet transform-based features for voice liveness detection

The need for Voice Liveness Detection (VLD) has emerged particularly for the security of Automatic Speaker Verification (ASV) systems. Existing Spoofed Speech Detection (SSD) systems rely on attack-specific approaches to detect spoofed speech. However, to safeguard ASV systems against all the kinds of spoofing attacks (known as well as unknown attacks), determining whether a speech is uttered live (genuine) or not, is important. To that effect, in this work, we propose the detection of pop noise using Morse wavelet for VLD task. Pop noise is a discriminative acoustic cue that is present in live speech and is absent/diminished in spoofed speech. It is captured by the microphone in the form of sudden bursts of air from a live speaker’s mouth due to the close proximity of the speaker with the microphone. To validate this hypothesis, we present an analysis of pop noise energy w.r.t. distance and found that it decreases exponentially with distance. Furthermore, pop noise is said to be present in very low frequency regions. To capture the pop noise effectively, we propose to exploit the excellent frequency resolution of Continuous Wavelet Transform (CWT) using Generalized Morse Wavelets (GMWs). GMWs are a superfamily of analytic wavelets. To that effect, in this work, we have analysed the suitability of GMWs for pop noise detection for VLD task using the POp noise COrpus (POCO). The wavelet parameters are fine-tuned according to the VLD task. Furthermore, the performance of VLD system is evaluated for various subband frequencies, and it is observed that the subband of 1 to 50Hz gives the best performance accuracy of 90.55% and 88.43% on the Dev and Eval sets, respectively. In addition, phoneme-based analysis shows the dependence of the performance of the VLD system on the type of phonemes in the utterances. It is shown that phonemes, such as plosives and fricatives show distinct pop noise as compared to other phonemes. Furthermore, the extension of the POCO dataset is used for experiments where simulated reverberation is added to spoofed signals, assuming the attacker (or the recording device) is positioned at various distances. This leads to the studying the effect of speaker-attacker distance. Similar to the previous results, it is observed that for the reverberated case too, the optimal frequency subband for VLD task is 1 to 50Hz, across all the distances. Furthermore, the proposed feature set is evaluated using three classifiers, namely, Convolutional Neural Network (CNN), Light CNN (LCNN), and Residual Neural Network (ResNet), for POCO dataset as well as reverberated POCO dataset. It is observed that CNN gives the highest accuracy of 88.43% on Eval set of the POCO dataset. Furthermore, the proposed features are also evaluated under the assumptions of two ideal scenarios — when the ASV system is strictly under attack, and when it is strictly not under attack. It is observed that the proposed Morse wavelet-based VLD system rejected 89% of the spoofed utterances, and accepted 88.30% of the genuine utterances.

Max Feature Map CNN with Support Vector Guided Softmax for Face Recognition

Face recognition has made significant progress because of advances in deep convolutional neural networks (CNNs) in addressing face verification in large amounts of data variation. When image data comes from different sources and devices, the identifiability of other classes and the presence of profile face data can lead to inaccurate and ambiguous classification because other classes lack discriminatory power. Furthermore, using a complex architecture with many deep convolutional layers can become very slow in the training process due to a huge amount of Random Access Memory (RAM) usage during the reverse pass of backpropagation. In this paper, we design a light CNN architecture that addresses these challenges. Specifically, we implemented Max-feature-map (MFM) into each convolutional layer to improve the accuracy and efficiency of the CNN. The strength of the support vector-guided SoftMax (SV-SoftMax) is also used in the proposed method to emphasize misclassified points and adaptively guide feature learning. Experimental results show that the 9-Layers CNN with MFM layer and SV-SoftMax outperform VGG-19 with 96.22% validation accuracy and the second rank below FaceNet tested on the same dataset with fewer parameters. Moreover, the model performed well on data that is obtained from various capture devices such as webcam, CCTVs, phone cameras, and DSLR cameras. The implications of this research could extend to scenarios requiring face recognition technology implementation with light size, such as surveillance and authentication systems

Deep learning-based automated defect classification in Electroluminescence images of solar panels

The rise of deep learning approaches has boosted the modern computing capabilities and paved the way to novel outstanding applications. Recently, deep neural networks (DNN) have caught an increasing attraction in the field of PV cells detection and fault classification. Convolutional neural networks (CNN) were proved as DNNs that provide more reliable results in diagnosing these faults. On the other hand, Electroluminescence imaging (EL) analysis has been recognized as a powerful tool for addressing PV models malfunctions. In this work, an effective fault detection and classification approach is developed using multi-scale CNN-based models using two scenarios: a) a transfer learning-based approach using two selected DNNs (ResNet18 and ShuffleNet), and b) an independent light-depth CNN (denoted as CNN-ILD). CNN-ILD is a light CNN based on the insertion of parallel convolutional branches with low-to-high kernel sizes. This construction allowed CNN-ILD to capture a wide variety of features ranging from elementary to specific textures and patterns. The public ELPV dataset has been adopted after proper preprocessing and class categorization. The experimental results have shown promising classification results (88.41%-98.05%) of PV cell defects in EL images. CNN-ILD has a privilege in terms of the computational power/time and stability over the other two proposed pretrained DNNs.

Differential convolutional network for noise mask estimation

With the advancements of speech synthesis technology, audio spoof detection systems have become vital for the security of automatic speaker verification systems. Many effective solutions have been offered for clean speech data. However, additive noise has a detrimental impact on the detection performance, as in many other speech related tasks. Noise mask is one of the methods proposed to increase the robustness against the additive noise. The purpose of the noise mask is to identify the time–frequency regions dominated by the noise signal. In this work, differential convolutional neural network is used to create noise masks. Differential convolution considers directional changes of the activations and generates new feature maps. Compared to the traditional convolutional network, a finer noise mask can be created with this method. Once the differential network for the noise masks is trained, its outputs are given to the spoof detection systems. Linear filterbank magnitudes are used as acoustic features for both noise masks and spoof detection. Therefore, the spoof detection systems have 2-channel inputs, i.e., linear filterbank magnitudes and its corresponding mask. Probabilistic linear discriminant analysis (PLDA) with x-vectors, emphasized channel attention, propagation and aggregation time delay neural network (ECAPA–TDNN), and light convolutional neural network (LCNN) followed by long short-term memory layers (LSTM) were used as classifiers. Three different noise types are used in both training and test stages, and two different noise types are used solely in the test stage, to stimulate seen and unseen conditions, respectively. Experiments conducted on the noisy version of ASVspoof 2015 challenge dataset showed that the LCNN–LSTM network with noise masks can achieve superior performance compared to other robust systems and can compete with the state-of-the-art. Considering the average of the known noise types, 2.67% equal error rate (EER) was observed. For the unknown noise types, 3.10% average EER was achieved. For the original (clean) ASVspoof 2015 data, the EER was 0.83%. Additionally, 2.6% EER was observed for logical access condition of ASVspoof 2019 data.

Intervention of light convolutional neural network in document survey form processing

Intervention of light convolutional neural network in document survey form processing

EFL-LCNN: Enhanced face localization augmented light convolutional neural network for human emotion recognition

EFL-LCNN: Enhanced face localization augmented light convolutional neural network for human emotion recognition

Light convolutional neural network by neural architecture search and model pruning for bearing fault diagnosis and remaining useful life prediction

Convolutional Neural Network (CNN) has been extensively used in bearing fault diagnosis and Remaining Useful Life (RUL) prediction. However, accompanied by CNN’s increasing performance is a deeper network structure and growing parameter size. This prevents it from being deployed in industrial applications with limited computation resources. To this end, this paper proposed a two-step method to build a cell-based light CNN by Neural Architecture Search (NAS) and weights-ranking-based model pruning. In the first step, a cell-based CNN was constructed with searched optimal cells and the number of stacking cells was limited to reduce the network size after influence analysis. To search for the optimal cells, a base CNN model with stacking cells was initially built, and Differentiable Architecture Search was adopted after continuous relaxation. In the second step, the connections in the built cell-based CNN were further reduced by weights-ranking-based pruning. Experiment data from the Case Western Reserve University was used for validation under the task of fault classification. Results showed that the CNN with only two cells achieved a test accuracy of 99.969% and kept at 99.968% even if 50% connections were removed. Furthermore, compared with base CNN, the parameter size of the 2-cells CNN was reduced from 9.677MB to 0.197MB. Finally, after minor revision, the network structure was adapted to achieve bearing RUL prediction and validated with the PRONOSTIA test data. Both tasks confirmed the feasibility and superiority of constructing a light cell-based CNN with NAS and pruning, which laid the potential to realize a light CNN in embedded systems.

Image Classification Based on Light Convolutional Neural Network Using Pulse Couple Neural Network.

Recently, most image classification studies solicit the intervention of convolutional neural networks because these DL-based classification methods generally outperform other methodologies with higher accuracy. However, this type of deep learning networks require many parameters and have a complex structure with multiple convolutional and pooling layers depending on the objective. These layers compute a large volume of data and it may impact the processing time and the performance. Therefore, this paper proposes a new method of image classification based on the light convolutional neural network. It consists of replacing the feature extraction layers of standard convolutional neural network with a single pulse coupled neural network by introducing the notion of foveation. This module provides the feature map of input image and the data compression using Discrete Wavelet Transform which is an optional step depending on the information quantity of this signature. The fully connected neural network, which has six hidden layers, classifies the image. With this technique, the computation time is reduced, and the network architecture is identical and simple independent of the type of dataset. The number of parameter is less than that in current research. The proposed method was validated with different dataset such as Caltech-101, Caltech-256, CIFAR-10, CIFAR-100, and ImageNet, and the accuracy reaches 92%, 90%, 99%, 94%, and 91%, respectively, which are better than the previous related works.

Novel Light Convolutional Neural Network for COVID Detection with Watershed Based Region Growing Segmentation.

A rapidly spreading epidemic, COVID-19 had a serious effect on millions and took many lives. Therefore, for individuals with COVID-19, early discovery is essential for halting the infection's progress. To quickly and accurately diagnose COVID-19, imaging modalities, including computed tomography (CT) scans and chest X-ray radiographs, are frequently employed. The potential of artificial intelligence (AI) approaches further explored the creation of automated and precise COVID-19 detection systems. Scientists widely use deep learning techniques to identify coronavirus infection in lung imaging. In our paper, we developed a novel light CNN model architecture with watershed-based region-growing segmentation on Chest X-rays. Both CT scans and X-ray radiographs were employed along with 5-fold cross-validation. Compared to earlier state-of-the-art models, our model is lighter and outperformed the previous methods by achieving a mean accuracy of 98.8% on X-ray images and 98.6% on CT scans, predicting the rate of 0.99% and 0.97% for PPV (Positive predicted Value) and NPV (Negative predicted Value) rate of 0.98% and 0.99%, respectively.

Light Deep Models for Cognitive Computing in Intelligent Transportation Systems

The paper proposes light convolutional neural network (CNN) models for the use of cognitive networking in an intelligent transportation system (ITS). There are two CNN models, one with 1D convolution and connectors, and the other with a tree-like structure. The 1D CNN model is deployed to process 1D temporal data such as the driver’s body temperature and electrocardiogram (ECG) data to measure emotion, while the deep tree CNN model is used to process image data obtained from car camera sensors. As the driver’s cognitive state can frequently change depending on the situation and location of the car, different edge controllers should handle the car sensors’ data within a short period of time. The tree-based deep learning model that can be branched and processed independently in the edge devices can be executed with less computation. This reduces the load and the time of the execution of the model. The light 1D CNN model has less learnable parameters, and hence can be executed in real-time. The cognitive state of a driver is measured by the facial emotion, body temperature, and ECG signal of the driver. The proposed is tested using a publicly available facial emotion database, and the accuracy and the information density are around 94-96% and 4.4, respectively.

Improved generalization performance of convolutional neural networks with LossDA

In recent years, convolutional neural networks (CNNs) have been used in many fields. Nowadays, CNNs have a high learning capability, and this learning capability is accompanied by a more complex model architecture. Complex model architectures allow CNNs to learn more data features, but such a learning process tends to reduce the training model’s ability to generalize to unknown data, and may be associated with problems of overfitting. Although many regularization methods have been proposed, such as data augmentation, batch normalization, and Dropout, research on improving generalization performance is still a common concern in the training process of robust CNNs. In this paper, we propose a dynamically controllable adjustment method, which we call LossDA, that embeds a disturbance variable in the fully-connected layer. The trend of this variable is kept consistent with the training loss, while the magnitude of the variable can be preset to adapt to the training process of different models. Through this dynamic adjustment, the training process of CNNs can be adaptively adjusted. The whole regularization process can improve the generalization performance of CNNs while helping to suppress overfitting. To evaluate this method, this paper conducts comparative experiments on MNIST, FashionMNIST, CIFAR-10, Cats_vs_Dogs, and miniImagenet datasets. The experimental results show that the method can improve the model performance of Light CNNs and Transfer CNNs (InceptionResNet, VGG19, ResNet50, and InceptionV3). The average maximum improvement in accuracy of Light CNNs is 4.62%, F1 is 3.99%, and Recall is 4.69%. The average maximum improvement accuracy of Transfer CNNs is 4.17%, F1 is 5.64%, and Recall is 4.05%.

Shoeprint Image Retrieval Based on Dual Knowledge Distillation for Public Security Internet of Things

In order to implement rapid retrieval of large-scale crime scene investigation shoeprint image (SPI) in the intelligent mobile terminal of Public Security Internet of Things (PSIoT), a novel dual knowledge distillation (DKD) network is designed by fusing spatial attention (SA) distillation and feature distillation to solve its problems of limited computing and storage capacity. First, ResNet50 was modified by adding a new SA module and hash layer as the teacher model, and a light convolutional neural network (CNN) with the same SA and hash layer is designed as the student model. Then, the attention distillation loss function is constructed to distill the SA knowledge in the teacher module to the student module to improve the ability of the convolutional layer at the front end of the student network to capture the underlying visual features of the SPI. Finally, the feature distillation loss function is constructed to distill the semantic knowledge in the teacher module to the student module to improve the ability of the student module to express high-level semantics of the SPI. we compare our method on two data sets of SPID and FID-300 with other state-of-the-art methods in the SPI retrieval domain. The experimental results show that our method can improve the SPI retrieval baseline by a large margin and better than other methods.

Enabling Gait Analysis in the Telemedicine Practice through Portable and Accurate 3D Human Pose Estimation

Human pose estimation (HPE) through deep learning-based software applications is a trend topic for markerless motion analysis. Thanks to the accuracy of the state-of-the-art technology, HPE could enable gait analysis in the telemedicine practice. On the other hand, delivering such a service at a distance requires the system to satisfy multiple and different constraints like accuracy, portability, real-time, and privacy compliance at the same time. Existing solutions either guarantee accuracy and real-time (e.g., the widespread OpenPose software on well-equipped computing platforms) or portability and data privacy (e.g., light convolutional neural networks on mobile phones). We propose a portable and low-cost platform that implements real-time and accurate 3D HPE through an embedded software on a low-power off-the-shelf computing device that guarantees privacy by default and by design. We present an extended evaluation of both accuracy and performance of the proposed solution conducted with a marker-based motion capture system (i.e., Vicon) as ground truth. The results show that the platform achieves real-time performance and high-accuracy with a deviation below the error tolerance when compared to the marker-based motion capture system (e.g., less than an error of 5∘ on the estimated knee flexion difference on the entire gait cycle and correlation 0.91<ρ<0.99). We provide a proof-of-concept study, showing that such portable technology, considering the limited discrepancies with respect to the marker-based motion capture system and its working tolerance, could be used for gait analysis at a distance without leading to different clinical interpretation.

PromoterLCNN: A Light CNN-Based Promoter Prediction and Classification Model.

Promoter identification is a fundamental step in understanding bacterial gene regulation mechanisms. However, accurate and fast classification of bacterial promoters continues to be challenging. New methods based on deep convolutional networks have been applied to identify and classify bacterial promoters recognized by sigma () factors and RNA polymerase subunits which increase affinity to specific DNA sequences to modulate transcription and respond to nutritional or environmental changes. This work presents a new multiclass promoter prediction model by using convolutional neural networks (CNNs), denoted as PromoterLCNN, which classifies Escherichia coli promoters into subclasses , , , , , and . We present a light, fast, and simple two-stage multiclass CNN architecture for promoter identification and classification. Training and testing were performed on a benchmark dataset, part of RegulonDB. Comparative performance of PromoterLCNN against other CNN-based classifiers using four parameters (Acc, Sn, Sp, MCC) resulted in similar or better performance than those that commonly use cascade architecture, reducing time by approximately 30–90% for training, prediction, and hyperparameter optimization without compromising classification quality.

Eye-referenced dynamic bounding box for face recognition using light convolutional neural network

Face recognition systems have imprinted its presence in many applications from offices to security to daily use as in personal digital gadgets. With many Face recognition systems in use, still there is scope for its performance improvement. The performance of such systems suffers due to presence of covariates like non-uniform illumination, pose deviations, occlusions, low resolution etc. Extracting the region of interest from input face images is very crucial steps in face recognition system. We proposed methodology to create a bounding box for every sample face image. This bounding box dynamically tries to fit the face image to cover the relevant face features useful for recognition purpose. We used a light CNN structure to perform the experiment on MUCT face dataset using three different methodologies for extracting region of interests from sample face images using bounding box. These are 1) min-max based ROI 2) OpenCV face alignment and 3) eye-reference based ROI. It is observed that the proposed eye-reference based face alignment method works better than conventional methods of min-max based ROI and OpenCV face alignment with considerable amount of improvement in the recognition accuracy. Our future work includes use of other structured datasets with various covariates for face recognition.