Center Loss Function Research Articles

Global-Local Multiple Granularity Learning for Cross-Modality Visible-Infrared Person Reidentification.

Cross-modality visible-infrared person reidentification (VI-ReID), which aims to retrieve pedestrian images captured by both visible and infrared cameras, is a challenging but essential task for smart surveillance systems. The huge barrier between visible and infrared images has led to the large cross-modality discrepancy and intraclass variations. Most existing VI-ReID methods tend to learn discriminative modality-sharable features based on either global or part-based representations, lacking effective optimization objectives. In this article, we propose a novel global-local multichannel (GLMC) network for VI-ReID, which can learn multigranularity representations based on both global and local features. The coarse- and fine-grained information can complement each other to form a more discriminative feature descriptor. Besides, we also propose a novel center loss function that aims to simultaneously improve the intraclass cross-modality similarity and enlarge the interclass discrepancy to explicitly handle the cross-modality discrepancy issue and avoid the model fluctuating problem. Experimental results on two public datasets have demonstrated the superiority of the proposed method compared with state-of-the-art approaches in terms of effectiveness.

Data-Driven Deepfake Forensics Model Based on Large-Scale Frequency and Noise Features

With the rapid development of deep learning and communication technology, the application of streaming media services and social software have gone deep into life. However, in the face of many uncertain factors in data dissemination, protecting privacy and security is particularly important. In order to solve the above problems, this study proposes a deep face forgery forensics method with frequency domain and noise features. In this method, DCT (i.e. Discrete Cosine Transform) transform is proposed to perceive the forgery trace features of different frequency bands in the frequency domain. At the same time, the spatial rich model (SRM) is used for guidance to enhance the traces of forged noise. Then, large-scale network and single center loss function are introduced to improve the forensics ability of the model. Experimental results on several databases such as faceforensics++, celeb DF and DFDC show that this method can effectively improve the accuracy of forensics.

Multi-Head Self-Attention-Enhanced Prototype Network with Contrastive–Center Loss for Few-Shot Relation Extraction

Few-shot relation extraction (FSRE) constitutes a critical task in natural language processing (NLP), involving learning relationship characteristics from limited instances to enable the accurate classification of new relations. The existing research primarily concentrates on using prototype networks for FSRE and enhancing their performance by incorporating external knowledge. However, these methods disregard the potential interactions among different prototype networks, and each prototype network can only learn and infer from its limited instances, which may limit the robustness and reliability of the prototype representations. To tackle the concerns outlined above, this paper introduces a novel prototype network called SACT (multi-head self-attention and contrastive-center loss), aimed at obtaining more comprehensive and precise interaction information from other prototype networks to bolster the reliability of the prototype network. Firstly, SACT employs a multi-head self-attention mechanism for capturing interaction information among different prototypes from traditional prototype networks, reducing the noise introduced by unknown categories with a small sample through information aggregation. Furthermore, SACT introduces a new loss function, the contrastive–center loss function, aimed at tightly clustering samples from a similar relationship category in the center of the feature space while dispersing samples from different relationship categories. Through extensive experiments on FSRE datasets, this paper demonstrates the outstanding performance of SACT, providing strong evidence for the effectiveness and practicality of SACT.

High-throughput soybean pods high-quality segmentation and seed-per-pod estimation for soybean plant breeding

Accurately identifying soybean pods is a crucial prerequisite for retrieving multi-phenotypic traits (such as number of pods per plant, number of seeds per pod, pod size, pod color, and pod shape). However, the traditional manual measurement approach for pod phenotype investigation is time-consuming and labor-intensive, particularly when counting the number of seeds per pod. Furthermore, existing instance segmentation methods designed for coarse-grained classification are inadequate for precise seed-per-pod estimation. To address these challenges, we modified an instance segmentation network for high-throughput soybean pods high-quality segmentation and accurate seed-per-pod estimation. We modified the classification branch of the instance segmentation network Mask Transfiner (Ke et al., 2022) by increasing the resolution of feature map of each Region of Interest (RoI) region, incorporating a dual attention mechanism, and leveraging a center loss function, named RefinePod. To overcome the limitation of scarce labeled data, we modified our synthesizing image method to automatedly generate fine-labeled multi-class soybean pods images. We then train RefinePod purely with these synthetic images. Subsequently, we evaluate the trained model on both synthetic and real test images. Experimental results demonstrate a significant improvement in the accuracy of seed-per-pod estimation achieved by RefinePod. Additionally, we conduct ablation experiments to analyze the individual contributions of each strategy employed in RefinePod. In summary, RefinePod achieves remarkable results in seed-per-pod estimation accuracy by integrating advanced techniques and leveraging synthetic images for training. Our findings highlight the potential of RefinePod for accelerating soybean phenotype investigation, enabling more efficient agricultural research and crop improvement initiatives.

An intelligent approach for anomaly detection in credit card data using bat optimization algorithm

As technology advances, many people are utilising credit cards to purchase their necessities, and the number of credit card scams is increasing tremendously. However, illegal card transactions have been on the rise, costing financial institutions millions of dollars each year. The development of efficient fraud detection techniques is critical in reducing these deficits, but it is difficult due to the extremely unbalanced nature of most credit card datasets. As compared to conventional fraud detection methods, the proposed method will help in automatically detecting the fraud, identifying hidden correlations in data and reduced time for verification process. This is achieved by selecting relevant and unique features by using Bat Optimization Algorithm (BOA). Next, balancing is performed in the highly imbalanced credit card fraud dataset using Synthetic Minority over-sampling technique (SMOTE). Then finally the CNN model for anomaly detection in credit card data is built using full center loss function to improve fraud detection performance and stability. The proposed model is tested with Kaggle dataset and yields around 99% accuracy.

Automated screening of computed tomography using weakly supervised anomaly detection.

Current artificial intelligence studies for supporting CT screening tasks depend on either supervised learning or detecting anomalies. However, the former involves a heavy annotation workload owing to requiring many slice-wise annotations (ground truth labels); the latter is promising, but while it reduces the annotation workload, it often suffers from lower performance. This study presents a novel weakly supervised anomaly detection (WSAD) algorithm trained based on scan-wise normal and anomalous annotations to provide better performance than conventional methods while reducing annotation workload. Based on surveillance video anomaly detection methodology, feature vectors representing each CT slice were trained on an AR-Net-based convolutional network using a dynamic multiple-instance learning loss and a center loss function. The following two publicly available CT datasets were retrospectively analyzed: the RSNA brain hemorrhage dataset (normal scans: 12,862; scans with intracranial hematoma: 8882) and COVID-CT set (normal scans: 282; scans with COVID-19: 95). Anomaly scores of each slice were successfully predicted despite inaccessibility to any slice-wise annotations. Slice-level area under the curve (AUC), sensitivity, specificity, and accuracy from the brain CT dataset were 0.89, 0.85, 0.78, and 0.79, respectively. The proposed method reduced the number of annotations in the brain dataset by 97.1% compared to an ordinary slice-level supervised learning method. This study demonstrated a significant annotation reduction in identifying anomalous CT slices compared to a supervised learning approach. The effectiveness of the proposed WSAD algorithm was verified through higher AUC than existing anomaly detection techniques.

LPI Radar Signals Modulation Recognition Based on ACDCA-ResNeXt

For low probability of intercept (LPI) radar waveform identification accuracy (ACC) problem at low Signal-to-Noise Ratios (SNRs), an approach based on time-frequency analysis (TFA) and Asymmetric Dilated Convolution Coordinate Attention Residual networks (ACDCA-ResNeXt) is proposed to recognize twelve kinds of LPI radar signals automatically. First, we apply Choi-Williams distribution (CWD), which shows superior performance at low SNRs, to transforming radar signals into time-frequency images (TFI). Then, in order to obtain the high-quality TFIs, a series of image processing techniques, including 2DWiener filtering, image cutting, and image resize, are used to remove the background noise and redundant frequency bands of the TFI and obtain a fixed-size gray scale image containing main morphological features of the TFI. Finally, the TFIs are input into ACDCA-ResNeXt network that can extract and learn deep features to recognize radar waveforms. Furthermore, a fusion loss function, which is composed of a soft-label smoothed cross entropy loss function and a center loss function, improves the generalization capability performance of network and achieves a better clustering effect. Experimental results demonstrate that, for twelve kinds of LPI radar waveforms, the overall recognition ACC of the proposed approach achieves 97.94% when SNR is -8 dB.

Unsupervised Domain Adaptive Person Re-Identification Method Based on Transformer

Person re-identification (ReID) is the problem of cross-camera target retrieval. The extraction of robust and discriminant features is the key factor in realizing the correct correlation of targets. A model based on convolutional neural networks (CNNs) can extract more robust image features. Still, it completes the extraction of images from local information to global information by continuously accumulating convolution layers. As a complex CNN, a vision transformer (ViT) captures global information from the beginning to extract more powerful features. This paper proposes an unsupervised domain adaptive person re-identification model (ViTReID) based on the vision transformer, taking the ViT model trained on ImageNet as the pre-training weight and a transformer encoder as the feature extraction network, which makes up for some defects of the CNN model. At the same time, the combined loss function of cross-entropy and triplet loss function combined with the center loss function is used to optimize the network; the person’s head is evaluated and trained as a local feature combined with the global feature of the whole body, focusing on the head, to enhance the head feature information. The experimental results show that ViTReID exceeds the baseline method (SSG) by 14% (Market1501 → MSMT17) in mean average precision (mAP). In MSMT17 → Market1501, ViTReID is 1.2% higher in rank-1 (R1) accuracy than a state-of-the-art method (SPCL); in PersonX → MSMT17, the mAP is 3.1% higher than that of the MMT-dbscan method, and in PersonX → Market1501, the mAP is 1.5% higher than that of the MMT-dbscan method.

A Novel Method for Ground-Based Cloud Image Classification Using Transformer

In recent years, convolutional neural networks (CNNs) have achieved competitive performance in the field of ground-based cloud image (GCI) classification. Proposed CNN-based methods can fully extract the local features of images. However, due to the locality of the convolution operation, they cannot well establish the long-range dependencies between the images, and thus they cannot extract the global features of images. Transformer has been applied to computer vision with great success due to its powerful global modeling capability. Inspired by it, we propose a Transformer-based GCI classification method that combines the advantages of the CNN and Transformer models. Firstly, the CNN model acts as a low-level feature extraction tool to generate local feature sequences of images. Then, the Transformer model is used to learn the global features of the images by efficiently extracting the long-range dependencies between the sequences. Finally, a linear classifier is used for GCI classification. In addition, we introduce a center loss function to address the problem of the simple cross-entropy loss not adequately supervising feature learning. Our method is evaluated on three commonly used datasets: ASGC, CCSN, and GCD. The experimental results show that the method achieves 94.24%, 92.73%, and 93.57% accuracy, respectively, outperforming other state-of-the-art methods. It proves that Transformer has great potential to be applied to GCI classification tasks.

Pedestrian Recognition System of Wild Scene Video Streaming Based on Convolutional Neural Network

Object detection aims at detecting the position of important objects and identifying them in still images or video frames. Pedestrians are the objects with most attention in images and videos. We study the detection and recognition technology for pedestrian objects of the wild scene video streams. Based on convolutional neural network, we design an online pedestrian recognition system. As non-rigid objects, people may exist in different poses in the image. Moreover, with the interference of complex background, lighting changes and occlusions, it will be very difficult to detect and recognize human objects. Following pedestrian detection in the images, we can conduct research on pedestrian tracking, pedestrian recognition and pedestrian behavior analysis. In addition, pedestrian detection technology also has great commercial value in video monitoring systems, intelligent traffic systems and other fields. In this paper, we propose a network model based on multiple loss functions, which combines the cross-entropy loss function with the improved monitoring signal center loss function to update the model parameters. Finally, the extracted features are more obvious, making the features of the same individual more compact. Compared to networks trained with only cross-entropy loss functions, our method significantly improves recognition accuracy and implements a pedestrian attribute recognition application based on Web technology. And compared with other similar applications, the recognition system we designed is cross-platform and lightweight. It mainly includes image acquisition, pedestrian attribute recognition and data storage modules, which is of great significance for the application of pedestrian recognition technology in the wild scenes.

An Extra-Contrast Affinity Network for Facial Expression Recognition in the Wild

Learning discriminative features for facial expression recognition (FER) in the wild is a challenging task due to the significant intra-class variations, inter-class similarities, and extreme class imbalances. In order to solve these issues, a contrastive-learning-based extra-contrast affinity network (ECAN) method is proposed. The ECAN consists of a feature processing network and two proposed loss functions, namely extra negative supervised contrastive loss (ENSC loss) and multi-view affinity loss (MVA loss). The feature processing network provides current and historical deep features to satisfy the necessary conditions for these loss functions. Specifically, the ENSC loss function simultaneously considers many positive samples and extra negative samples from other minibatches to maximize intra-class similarity and the inter-class separation of deep features, while also automatically turning the attention of the model to majority and minority classes to alleviate the class imbalance issue. The MVA loss function improves upon the center loss function by leveraging additional deep feature groups from other minibatches to dynamically learn more accurate class centers and further enhance the intra-class compactness of deep features. The numerical results obtained using two public wild FER datasets (RAFDB and FER2013) indicate that the proposed method outperforms most state-of-the-art models in FER.

An automatic residual-constrained and clustering-boosting architecture for differentiated heartbeat classification

ObjectiveThis paper proposes a residual-constrained and clustering-boosting architecture for automatic patient-specific heartbeat classification. MethodsHeartbeats are segmented from long-term ECG recordings with a customized method to make the heartbeats selectively focus on rhythm information and increase the inter-class difference. The Residual-Squeeze-and-Excitation (Res-SE) block is used in the proposed network to reduce the loss of shallow features and emphasize the important features. The center loss function is applied to constrain the output feature vectors and increase the intra-class compactness of the learned features. Then the output features are concatenated with RR-interval features to calculate the classification results. Finally, the classifier predictions are fine-tuned with clustering-boosting strategy. ResultsThis architecture is verified on MIT-BIH Arrhythmia database (MIT-DB) and MIT Supraventricular Arrhythmia database (MIT-SVDB). Our method attains F1 scores of 0.96 and 0.90 for ventricular ectopic beat (VEB) and supraventricular ectopic beat (SVEB) respectively, and has good generalization ability. ConclusionThe proposed method, combining supervised learning with unsupervised learning, provides a new architecture for heartbeat classification. It has shown excellent performance on multi-class heartbeat classification. SignificanceThe proposed method has the potential to be applied for clinical ECG diagnosis.

A Music Emotion Classification Model Based on the Improved Convolutional Neural Network.

Aiming at the problems of music emotion classification, a music emotion recognition method based on the convolutional neural network is proposed. First, the mel-frequency cepstral coefficient (MFCC) and residual phase (RP) are weighted and combined to extract the audio low-level features of music, so as to improve the efficiency of data mining. Then, the spectrogram is input into the convolutional recurrent neural network (CRNN) to extract the time-domain features, frequency-domain features, and sequence features of audio. At the same time, the low-level features of audio are input into the bidirectional long short-term memory (Bi-LSTM) network to further obtain the sequence information of audio features. Finally, the two parts of features are fused and input into the softmax classification function with the center loss function to achieve the recognition of four music emotions. The experimental results based on the emotion music dataset show that the recognition accuracy of the proposed method is 92.06%, and the value of the loss function is about 0.98, both of which are better than other methods. The proposed method provides a new feasible idea for the development of music emotion recognition.

Research on Face Intelligent Perception Technology Integrating Deep Learning under Different Illumination Intensities

Aiming at the problem of face recognition under different illumination intensities combined with deep learning algorithm, this research designs a new loss function, i-center loss function and integrates the structure of migration learning algorithm on the basis of LeNets++ deep learning network. The face image data set labled faces in the wild with different illumination intensities and the image data set of supermarket monitoring system are used to train and test the improved LeNets++ deep learning network based on softmax, center and i-center loss function, and a variety of common image recognition networks. The calculation results show that although the amount of data required for the training of LeNets++ deep learning network is much larger than other networks selected in the study, when the loss function is changed to i-center, the accuracy of face image recognition under different light intensities is significantly improved, reaching 99.65%. In the supermarket data set, the maximum face recognition rate of the algorithm using i-center loss function is 99.07%, which is 0.21% and 0.6% higher than that of using center softmax and softmax loss function, respectively. Therefore, experiments show that the improved deep learning neural network based on i-center loss function can improve the effect of face recognition under different illumination intensities.

An effective crack position diagnosis method for the hollow shaft rotor system based on the convolutional neural network and deep metric learning

In recent years, the crack fault is one of the most common faults in the rotor system and it is still a challenge for crack position diagnosis in the hollow shaft rotor system. In this paper, a method based on the Convolutional Neural Network and deep metric learning (CNN-C) is proposed to effectively identify the crack position for a hollow shaft rotor system. Center-loss function is used to enhance the performance of neural network. Main contributions include: Firstly, the dynamic response of the dual-disks hollow shaft rotor system is obtained. The analysis results show that the crack will cause super-harmonic resonance, and the peak value of it is closely related to the position and depth of the crack. In addition, the amplitude near the non-resonant region also has relationship with the crack parameters. Secondly, we proposed an effective crack position diagnosis method which has the highest 99.04% recognition accuracy compared with other algorithms. Then, the influence of penalty factor on CNN-C performance is analyzed, which shows that too high penalty factor will lead to the decline of the neural network performance. Finally, the feature vectors are visualized via t-distributed Stochastic Neighbor Embedding (t-SNE). Naive Bayes classifier (NB) and K-Nearest Neighbor algorithm (KNN) are used to verify the validity of the feature vectors extracted by CNN-C. The results show that NB and KNN have more regular decision boundaries and higher recognition accuracy on the feature vectors data set extracted by CNN-C, indicating that the feature vectors extracted by CNN-C have great intra-class compactness and inter-class separability.

An Intra-Pulse Modulation Type Recognition Algorithm for Radar Signals Based on the Improved Residual Network

For purpose of solve the problem of poor discrimination and robustness of intra-pulse signal features extracted by the traditional methods, this paper proposes a radar signal intra-pulse modulation type recognition algorithm based on the improved residual network. Firstly, one-dimensional time-domain radar signal is converted into two-dimensional time-frequency image by Smoothing Pseudo Wigner-Ville Distribution; Then the time-frequency image is preprocessed; ResNet-50 network is chosen as the framework. In order to retain the feature map information as much as possible, the convolution kernel is increased in the residual module. The cross entropy loss function and the center loss function are used as the loss function to speed up the convergence of the network. The improved residual network is used to realize the intra-pulse modulation type recognition of radar signal. The simulation experiments show that when the SNR is -14dB, the overall average recognition accuracy of the improved algorithm for eight kinds of radar signals (CM, LFM, NLFM, BLFM, BPSK, QPSK, OPSK, LFM+BPSK) can reach 97.29%, which shows the effectiveness.

Joint feature embedding learning and correlation filters for aircraft tracking with infrared imagery

Infrared object tracking is a key technology for infrared imaging guidance. Blurred imaging, strong ego-motion and frequent occlusion make it difficult to maintain robust tracking. We observe that the features trained on ImageNet are not suitable for aircraft tracking with infrared imagery. In addition, for deep feature-based tracking, the main computational burden comes from the feedforward pass through the pretrained deep network. To this end, we present an airborne infrared target tracking algorithm that employs feature embedding learning and correlation filters to obtain improved performance. We develop a shallow network and a contrastive center loss function to learn the prototypical representation of the aircraft in the embedding space. The feature embedding module is lightweight and integrated into the efficient convolution operator framework for aircraft tracking. Finally, to demonstrate the effectiveness of our tracking algorithm, we conduct extensive experiments on airborne infrared imagery and benchmark trackers.

A Novel Face Super-Resolution Method Based on Parallel Imaging and OpenVINO

Face image super-resolution refers to recovering a high-resolution face image from a low-resolution one. In recent years, due to the breakthrough progress of deep representation learning for super-resolution, the study of face super-resolution has become one of the hot topics in the field of super-resolution. However, the performance of these deep learning-based approaches highly relies on the scale of training samples and is limited in efficiency in real-time applications. To address these issues, in this work, we introduce a novel method based on the parallel imaging theory and OpenVINO. In particular, inspired by the methodology of learning-by-synthesis in parallel imaging, we propose to learn from the combination of virtual and real face images. In addition, we introduce a center loss function borrowed from the deep model to enhance the robustness of our model and propose to apply OpenVINO to speed up the inference. To the best of our knowledge, it is the first time to tackle the problem of face super-resolution based on parallel imaging methodology and OpenVINO. Extensive experimental results and comparisons on the publicly available LFW, WebCaricature, and FERET datasets demonstrate the effectiveness and efficiency of the proposed method.

Learn class hierarchy using convolutional neural networks

A large amount of research on Convolutional Neural Networks (CNN) has focused on flat Classification in the multi-class domain. In the real world, many problems are naturally expressed as hierarchical classification problems, in which the classes to be predicted are organized in a hierarchy of classes. In this paper, we propose a new architecture for hierarchical classification, introducing a stack of deep linear layers using cross-entropy loss functions combined to a center loss function. The proposed architecture can extend any neural network model and simultaneously optimizes loss functions to discover local hierarchical class relationships and a loss function to discover global information from the whole class hierarchy while penalizing class hierarchy violations. We experimentally show that our hierarchical classifier presents advantages to the traditional classification approaches finding application in computer vision tasks. The same approach can also be applied to some CNN for text classification.

Underwater Acoustic Target Recognition with a Residual Network and the Optimized Feature Extraction Method

Underwater Acoustic Target Recognition (UATR) remains one of the most challenging tasks in underwater signal processing due to the lack of labeled data acquisition, the impact of the time-space varying intrinsic characteristics, and the interference from other noise sources. Although some deep learning methods have been proven to achieve state-of-the-art accuracy, the accuracy of the recognition task can be improved by designing a Residual Network and optimizing feature extraction. To give a more comprehensive representation of the underwater acoustic signal, we first propose the three-dimensional fusion features along with the data augment strategy of SpecAugment. Afterward, an 18-layer Residual Network (ResNet18), which contains the center loss function with the embedding layer, is designed to train the aggregated features with an adaptable learning rate. The recognition experiments are conducted on the ship-radiated noise dataset from a real environment, and the accuracy results of 94.3% indicate that the proposed method is appropriate for underwater acoustic recognition problems and sufficiently surpasses other classification methods.