Intermediate Feature Maps Research Articles

Enhancing brain tumor classification through ensemble attention mechanism.

Brain tumors pose a serious threat to public health, impacting thousands of individuals directly or indirectly worldwide. Timely and accurate detection of these tumors is crucial for effective treatment and enhancing the quality of patients' lives. The widely used brain imaging technique is magnetic resonance imaging, the precise identification of brain tumors in MRI images is challenging due to the diverse anatomical structures. This paper introduces an innovative approach known as the ensemble attention mechanism to address this challenge. Initially, the approach uses two networks to extract intermediate- and final-level feature maps from MobileNetV3 and EfficientNetB7. This assists in gathering the relevant feature maps from the different models at different levels. Then, the technique incorporates a co-attention mechanism into the intermediate and final feature map levels on both networks and ensembles them. This directs attention to certain regions to extract global-level features at different levels. Ensemble of attentive feature maps enabling the precise detection of various feature patterns within brain tumor images at both model, local, and global levels. This leads to an improvement in the classification process. The proposed system was evaluated on the Figshare dataset and achieved an accuracy of 98.94%, and 98.48% for the BraTS 2019 dataset which is superior to other methods. Thus, it is robust and suitable for brain tumor detection in healthcare systems.

A simplified adversarial architecture for cross-subject silent speech recognition using electromyography

Objective. The decline in the performance of electromyography (EMG)-based silent speech recognition is widely attributed to disparities in speech patterns, articulation habits, and individual physiology among speakers. Feature alignment by learning a discriminative network that resolves domain offsets across speakers is an effective method to address this problem. The prevailing adversarial network with a branching discriminator specializing in domain discrimination renders insufficiently direct contribution to categorical predictions of the classifier.Approach. To this end, we propose a simplified discrepancy-based adversarial network with a streamlined end-to-end structure for EMG-based cross-subject silent speech recognition. Highly aligned features across subjects are obtained by introducing a Nuclear-norm Wasserstein discrepancy metric on the back end of the classification network, which could be utilized for both classification and domain discrimination. Given the low-level and implicitly noisy nature of myoelectric signals, we devise a cascaded adaptive rectification network as the front-end feature extraction network, adaptively reshaping the intermediate feature map with automatically learnable channel-wise thresholds. The resulting features effectively filter out domain-specific information between subjects while retaining domain-invariant features critical for cross-subject recognition.Main results. A series of sentence-level classification experiments with 100 Chinese sentences demonstrate the efficacy of our method, achieving an average accuracy of 89.46% tested on 40 new subjects by training with data from 60 subjects. Especially, our method achieves a remarkable 10.07% improvement compared to the state-of-the-art model when tested on 10 new subjects with 20 subjects employed for training, surpassing its result even with three times training subjects.Significance. Our study demonstrates an improved classification performance of the proposed adversarial architecture using cross-subject myoelectric signals, providing a promising prospect for EMG-based speech interactive application.

Gradient-Based Instance-Specific Visual Explanations for Object Specification and Object Discrimination.

We propose the gradient-weighted Object Detector Activation Maps (ODAM), a visual explanation technique for interpreting the predictions of object detectors. Utilizing the gradients of detector targets flowing into the intermediate feature maps, ODAM produces heat maps that show the influence of regions on the detector's decision for each predicted attribute. Compared to previous works on classification activation maps (CAM), ODAM generates instance-specific explanations rather than class-specific ones. We show that ODAM is applicable to one-stage, two-stage, and transformer-based detectors with different types of detector backbones and heads, and produces higher-quality visual explanations than the state-of-the-art in terms of both effectiveness and efficiency. We discuss two explanation tasks for object detection: 1) object specification: what is the important region for the prediction? 2) object discrimination: which object is detected? Aiming at these two aspects, we present a detailed analysis of the visual explanations of detectors and carry out extensive experiments to validate the effectiveness of the proposed ODAM. Furthermore, we investigate user trust on the explanation maps, how well the visual explanations of object detectors agrees with human explanations, as measured through human eye gaze, and whether this agreement is related with user trust. Finally, we also propose two applications, ODAM-KD and ODAM-NMS, based on these two abilities of ODAM. ODAM-KD utilizes the object specification of ODAM to generate top-down attention for key predictions and instruct the knowledge distillation of object detection. ODAM-NMS considers the location of the model's explanation for each prediction to distinguish the duplicate detected objects. A training scheme, ODAM-Train, is proposed to improve the quality on object discrimination, and help with ODAM-NMS.

A Novel Dual Attention Approach for DNN Based Automated Diabetic Retinopathy Grading

ABSTRACTDiabetic retinopathy (DR) poses a serious threat to vision, emphasising the need for early detection. Manual analysis of fundus images, though common, is error‐prone and time‐intensive. Existing automated diagnostic methods lack precision, particularly in the early stages of DR. This paper introduces the Soft Convolutional Block Attention Module‐based Network (Soft‐CBAMNet), a deep learning network designed for severity detection, which features Soft‐CBAM attention to capture complex features from fundus images. The proposed network integrates both the convolutional block attention module (CBAM) and the soft‐attention components, ensuring simultaneous processing of input features. Following this, attention maps undergo a max‐pooling operation, and refined features are concatenated before passing through a dropout layer with a dropout rate of 50%. Experimental results on the APTOS dataset demonstrate the superior performance of Soft‐CBAMNet, achieving an accuracy of 85.4% in multiclass DR grading. The proposed architecture has shown strong robustness and general feature learning capability, achieving a mean AUC of 0.81 on the IDRID dataset. Soft‐CBAMNet's dynamic feature extraction capability across all classes is further justified by the inspection of intermediate feature maps. The model excels in identifying all stages of DR with increased precision, surpassing contemporary approaches. Soft‐CBAMNet presents a significant advancement in DR diagnosis, offering improved accuracy and efficiency for timely intervention.

Improving the transferability of adversarial examples through black-box feature attacks

Deep neural networks (DNNs) are vulnerable and susceptible to imperceptible perturbations. Adversarial examples become more and more popular. Black-box attacks are considered to be the most realistic scenario. Currently, transfer-based black-box attacks show excellent performance. However, transfer-based black-box attacks all require an agent model of the attack, which we call the source model. This leads to the existing transfer-based attacks limited by the features focused on the source model, which creates a bottleneck in improving the transferability of adversarial examples. In order to solve this problem, we propose an attack that mainly targets features that are insensitive to the source model, which we call the black-box feature attack. Specifically, we categorize the features of the image into white-box features and black-box features. The white-box features are source model-sensitive features and the black-box features are source model insensitive features. White-box features are only specific to the source model, while black-box features are more generalized for unknown models. By destroying the image white-box features, the fitted image is obtained and the model intermediate layer feature map is extracted. Afterward, the fitting gradient is found for the fitted images with different fitting degrees. We construct loss functions based on the obtained fitting gradients and feature maps to guide the attacks to better destroy the black-box features of the images. Extensive experiments demonstrate that our methods have higher transferability compared to state-of-the-art methods, which achieve more than 90% of transferability under the normal model. It is also significantly better than other methods on adversarially trained models. Even in the white-box setting, our attack has the best performance.

Style2Talker: High-Resolution Talking Head Generation with Emotion Style and Art Style

Although automatically animating audio-driven talking heads has recently received growing interest, previous efforts have mainly concentrated on achieving lip synchronization with the audio, neglecting two crucial elements for generating expressive videos: emotion style and art style. In this paper, we present an innovative audio-driven talking face generation method called Style2Talker. It involves two stylized stages, namely Style-E and Style-A, which integrate text-controlled emotion style and picture-controlled art style into the final output. In order to prepare the scarce emotional text descriptions corresponding to the videos, we propose a labor-free paradigm that employs large-scale pretrained models to automatically annotate emotional text labels for existing audio-visual datasets. Incorporating the synthetic emotion texts, the Style-E stage utilizes a large-scale CLIP model to extract emotion representations, which are combined with the audio, serving as the condition for an efficient latent diffusion model designed to produce emotional motion coefficients of a 3DMM model. Moving on to the Style-A stage, we develop a coefficient-driven motion generator and an art-specific style path embedded in the well-known StyleGAN. This allows us to synthesize high-resolution artistically stylized talking head videos using the generated emotional motion coefficients and an art style source picture. Moreover, to better preserve image details and avoid artifacts, we provide StyleGAN with the multi-scale content features extracted from the identity image and refine its intermediate feature maps by the designed content encoder and refinement network, respectively. Extensive experimental results demonstrate our method outperforms existing state-of-the-art methods in terms of audio-lip synchronization and performance of both emotion style and art style.

A Lightweight Channel and Time Attention Enhanced 1D CNN Model for Environmental Sound Classification

One dimension convolutional neural networks (1D CNN) that directly take raw waveforms as input has less competition than 2D CNN recognizing environmental sound. In order to overcome its disadvantages, we propose a novel lightweight 1D CNN structure by employing attention mechanism, which has significant improvement in both accuracy and computational complexity. Concretely, (1) two attention modules are constructed along channel and time dimension separately, and combined to give an intermediate feature map, which focus on key frequency band and semantically related time frame information. (2) Without increasing training overhead, snapshot ensemble is employed to further improve performance. Results from two benchmarking datasets (UrbanSound8k, ESC-10) demonstrated that: by employing attention mechanism, our model outperforms all of the previously reported 1D CNN approaches in accuracy with less parameters. Meanwhile with improved performance gain, the proposed model is superior than most of the existing spectral-based 2D CNN approaches and competitive with SOTA performance, while with orders of magnitude parameters fewer. Overall, it indicates our model is compact and has good potential in practical resource-limited applications, such as sound recognition on embedded platform.

PADVG: A Simple Baseline of Active Protection for Audio-Driven Video Generation

Over the past few years, deep generative models have significantly evolved, enabling the synthesis of realistic content and also bringing security concerns of illegal misuse. Therefore, active protection for generative models has been proposed recently, aiming to generate samples with hidden messages for future identification while preserving the original generating performance. However, existing active protection methods are specifically designed for generative adversarial networks (GANs), restricted to handling unconditional image generation. We observe that they get limited identification performance and visual quality when handling audio-driven video generation conditioned on target audio and source input to drive video generation with consistent context, e.g., identity and movement, between frame sequences. To address this issue, we introduce a simple yet effective active P rotection framework for A udio- D riven V ideo G eneration, named PADVG. To be specific, we present a novel frame-shared embedding module in which messages to hide are first transformed into frame-shared message coefficients. Then, these coefficients are assembled with the intermediate feature maps of video generators at multiple feature levels to generate the embedded video frames. Besides, PADVG further considers two visual consistent losses: (i) intra-frame loss is utilized to keep the visual consistency with different hidden messages; (ii) inter-frame loss is used to preserve the visual consistency across different video frames. Moreover, we also propose an auxiliary denoising training strategy through perturbing the assembled features by learnable pixel-level noise to improve identification performance, while enhancing robustness against real-world disturbances. Extensive experiments demonstrate that our proposed PADVG for audio-driven video generation can effectively identify the generated videos and achieve high visual quality.

Approximating Intermediate Feature Maps of Self-Supervised Convolution Neural Network to Learn Hard Positive Representations in Chest Radiography.

Recent advances in contrastive learning have significantly improved the performance of deep learning models. In contrastive learning of medical images, dealing with positive representation is sometimes difficult because some strong augmentation techniques can disrupt contrastive learning owing to the subtle differences between other standardized CXRs compared to augmented positive pairs; therefore, additional efforts are required. In this study, we propose intermediate feature approximation (IFA) loss, which improves the performance of contrastive convolutional neural networks by focusing more on positive representations of CXRs without additional augmentations. The IFA loss encourages the feature maps of a query image and its positive pair to resemble each other by maximizing the cosine similarity between the intermediate feature outputs of the original data and the positive pairs. Therefore, we used the InfoNCE loss, which is commonly used loss to address negative representations, and the IFA loss, which addresses positive representations, together to improve the contrastive network. We evaluated the performance of the network using various downstream tasks, including classification, object detection, and a generative adversarial network (GAN) inversion task. The downstream task results demonstrated that IFA loss can improve the performance of effectively overcoming data imbalance and data scarcity; furthermore, it can serve as a perceptual loss encoder for GAN inversion. In addition, we have made our model publicly available to facilitate access and encourage further research and collaboration in the field.

Efficient Crowd Counting via Dual Knowledge Distillation.

Most researchers focus on designing accurate crowd counting models with heavy parameters and computations but ignore the resource burden during the model deployment. A real-world scenario demands an efficient counting model with low-latency and high-performance. Knowledge distillation provides an elegant way to transfer knowledge from a complicated teacher model to a compact student model while maintaining accuracy. However, the student model receives the wrong guidance with the supervision of the teacher model due to the inaccurate information understood by the teacher in some cases. In this paper, we propose a dual-knowledge distillation (DKD) framework, which aims to reduce the side effects of the teacher model and transfer hierarchical knowledge to obtain a more efficient counting model. First, the student model is initialized with global information transferred by the teacher model via adaptive perspectives. Then, the self-knowledge distillation forces the student model to learn the knowledge by itself, based on intermediate feature maps and target map. Specifically, the optimal transport distance is utilized to measure the difference of feature maps between the teacher and the student to perform the distribution alignment of the counting area. Extensive experiments are conducted on four challenging datasets, demonstrating the superiority of DKD. When there are only approximately 6% of the parameters and computations from the original models, the student model achieves a faster and more accurate counting performance as the teacher model even surpasses it.

Attentional Feature Erase: Towards task-wise transferable adversarial attack on cloud vision APIs

Recent works have shown that adversarial examples (AEs) can attack and successfully transfer across various neural networks, highlighting the potential danger they pose. However, current approaches that focus on task-specific loss functions may not be as effective across different tasks. Additionally, the use of cloud APIs in practice, which often involve combining multiple tasks, also weakens the effectiveness of existing attacks. To address these issues, we propose a method called Attentional Feature Erase, which is a task-agnostic attack with improved cross-task transferability and effectiveness on computer vision-based cloud APIs. We view the transferability of AEs as a latent contribution for each layer of deep neural networks. By focusing on the intermediate layers of model backbones and reducing high-value features in each intermediate feature map, we are able to maximize the attack performance. Additionally, to better aggregate the gradients and generate adversarial perturbations during backward propagation, Transferability Regularizer is proposed to calculate the attention heatmap for each intermediate feature map and systematically combine the gradients. Comprehensive set of experiments on the Google Cloud Vision APIs and public available datasets (i.e. ImageNet, PASCAL VOC and MS COCO) show that the proposed AFE attack is more effective and has better transferability compared to the state-of-the-art baselines.

Fastensor: Optimise the Tensor I/O Path from SSD to GPU for Deep Learning Training

In recent years, benefiting from the increase in model size and complexity, deep learning has achieved tremendous success in computer vision (CV) and (NLP). Training deep learning models using accelerators such as GPUs often requires much iterative data to be transferred from NVMe SSD to GPU memory. Much recent work has focused on data transfer during the pre-processing phase and has introduced techniques such as multiprocessing and GPU Direct Storage (GDS) to accelerate it. However, tensor data during training (such as Checkpoints, logs, and intermediate feature maps), which is also time-consuming, is often transferred using traditional serial, long-I/O-path transfer methods. In this article, based on GDS technology, we built Fastensor, an efficient tool for tensor data transfer between the NVMe SSDs and GPUs. To achieve higher tensor data I/O throughput, we optimized the traditional data I/O process. We also proposed a data and runtime context-aware tensor I/O algorithm. Fastensor can select the most suitable data transfer tool for the current tensor from a candidate set of tools during model training. The optimal tool is derived from a dictionary generated by our adaptive exploration algorithm in the first few training iterations. We used Fastensor’s unified interface to test the read/write bandwidth and energy consumption of different transfer tools for different sizes of tensor blocks. We found that the execution efficiency of different tensor transfer tools is related to both the tensor block size and the runtime context. We then deployed Fastensor in the widely applicable Pytorch deep learning framework. We showed that Fastensor could perform superior in typical scenarios of model parameter saving and intermediate feature map transfer with the same hardware configuration. Fastensor achieves a 5.37x read performance improvement compared to torch.save () when used for model parameter saving. When used for intermediate feature map transfer, Fastensor can increase the supported training batch size by 20x, while the total read and write speed is increased by 2.96x compared to the torch I/O API.

Multi-style spatial attention module for cortical cataract classification in AS-OCT image with supervised contrastive learning

Background and Objective: Precise cortical cataract (CC) classification plays a significant role in early cataract intervention and surgery. Anterior segment optical coherence tomography (AS-OCT) images have shown excellent potential in cataract diagnosis. However, due to the complex opacity distributions of CC, automatic AS-OCT-based CC classification has been rarely studied. In this paper, we aim to explore the opacity distribution characteristics of CC as clinical priori to enhance the representational capability of deep convolutional neural networks (CNNs) in CC classification tasks.Methods: We propose a novel architectural unit, Multi-style Spatial Attention module (MSSA), which recalibrates intermediate feature maps by exploiting diverse clinical contexts. MSSA first extracts the clinical style context features with Group-wise Style Pooling (GSP), then refines the clinical style context features with Local Transform (LT), and finally executes group-wise feature map recalibration via Style Feature Recalibration (SFR). MSSA can be easily integrated into modern CNNs with negligible overhead.Results: The extensive experiments on a CASIA2 AS-OCT dataset and two public ophthalmic datasets demonstrate the superiority of MSSA over state-of-the-art attention methods. The visualization analysis and ablation study are conducted to improve the explainability of MSSA in the decision-making process.Conclusions: Our proposed MSSANet utilized the opacity distribution characteristics of CC to enhance the representational power and explainability of deep convolutional neural network (CNN) and improve the CC classification performance. Our proposed method has the potential in the early clinical CC diagnosis.

The Identification of Fritillaria Species Using Hyperspectral Imaging with Enhanced One-Dimensional Convolutional Neural Networks via Attention Mechanism.

Combining deep learning and hyperspectral imaging (HSI) has proven to be an effective approach in the quality control of medicinal and edible plants. Nonetheless, hyperspectral data contains redundant information and highly correlated characteristic bands, which can adversely impact sample identification. To address this issue, we proposed an enhanced one-dimensional convolutional neural network (1DCNN) with an attention mechanism. Given an intermediate feature map, two attention modules are constructed along two separate dimensions, channel and spectral, and then combined to enhance relevant features and to suppress irrelevant ones. Validated by Fritillaria datasets, the results demonstrate that an attention-enhanced 1DCNN model outperforms several machine learning algorithms and shows consistent improvements over a vanilla 1DCNN. Notably under VNIR and SWIR lenses, the model obtained 98.97% and 99.35% for binary classification between Fritillariae Cirrhosae Bulbus (FCB) and other non-FCB species, respectively. Additionally, it still achieved an extraordinary accuracy of 97.64% and 98.39% for eight-category classification among Fritillaria species. This study demonstrated the application of HSI with artificial intelligence can serve as a reliable, efficient, and non-destructive quality control method for authenticating Fritillaria species. Moreover, our findings also illustrated the great potential of the attention mechanism in enhancing the performance of the vanilla 1DCNN method, providing reference for other HSI-related quality controls of plants with medicinal and edible uses.

Module of Axis-based Nexus Attention for weakly supervised object localization

Weakly supervised object localization tasks remain challenging to identify and segment an entire object rather than only discriminative parts of the object. To tackle this problem, corruption-based approaches have been devised, which involve the training of non-discriminative regions by corrupting (e.g., erasing) the input images or intermediate feature maps. However, this approach requires an additional hyperparameter, the corrupting threshold, to determine the degree of corruption and can unfavorably disrupt training. It also tends to localize object regions coarsely. In this paper, we propose a novel approach, Module of Axis-based Nexus Attention (MoANA), which helps to adaptively activate less discriminative regions along with the class-discriminative regions without an additional hyperparameter, and elaborately localizes an entire object. Specifically, MoANA consists of three mechanisms (1) triple-view attentions representation, (2) attentions expansion, and (3) features calibration mechanism. Unlike other attention-based methods that train a coarse attention map with the same values across elements in feature maps, MoANA trains fine-grained values in an attention map by assigning different attention values to each element. We validated MoANA by comparing it with various methods. We also analyzed the effect of each component in MoANA and visualized attention maps to provide insights into the calibration.

MistNet: A superior edge-cloud privacy-preserving training framework with one-shot communication

Classical federated learning methods aggregate decentralized data from different devices into a central location for efficient training. However, these approaches raise significant concerns regarding data privacy and requires extensive communication between edge devices and the cloud, leading to high communication costs. To conquer these challenges, we propose an edge-cloud privacy-preserving model training system that enables the cloud and edge devices to collaboratively perform deep neural network(DNN) training, namely MistNet, which has the following advantages. (1) Different from previous works that required frequent communication between the edge and cloud, MistNet requires only one-shot communication. (2) MistNet introduces the rigorous local differential privacy (LDP) technique to the intermediate feature maps, guaranteeing that the user’s local data and model parameters remain undisclosed. (3) In MistNet, the feature extractor is transferred from pre-trained models that are designed for similar application domains. This feature extractor remains fixed during training, thereby eliminating the requirement to synchronize feature extractors across different devices. Furthermore, we first design an object detection algorithm based on Yolov5 with the MistNet framework. Finally, the extensive experiments results on multiple models and datasets demonstrate that by choosing an appropriate partition layer and privacy budget, MistNet achieves lower communication and faster convergence, as well as acceptable model utility while greatly reducing privacy leakage from the released intermediate features. Our framework is primarily designed for image data and has already been applied to Plato. You can find the source code on GitHub page at https://github.com/TL-System/plato.

LHFNet: A Fast and Accurate Object Detector Based on CenterNet

This paper proposes LHFNet, an improved object detection method based on CenterNet, which achieves a better speed-accuracy trade-off. The main contributions of this paper are as follows: We design a lightweight Hourglass network to reduce the number of parameters and computations of CenterNet, thus improving the detection speed. We introduce an intermediate feature map fusion module to enhance the feature extraction capability of Hourglass, which compensates for the possible performance degradation caused by the network simplification. We apply an attention module to the final heat map to improve the image feature representation. We design and optimize a series of loss functions to train LHFNet effectively. Through the research in this paper, we demonstrate that the feature map fusion and the attention module greatly improve the performance of LHFNet. We evaluate our method on the COCO dataset and achieve a high accuracy at a fast speed (Tesla V100), which outperforms other object detection methods in terms of both accuracy and efficiency.

Fuel consumption prediction for pre-departure flights using attention-based multi-modal fusion

Improper fuel loading decision results in carrying excessive dead weight during flight operation, which will burden the airline operation cost and cause extra waste emission. Existing works mainly focused on the post-event fuel consumption based on flight trajectory. In this work, a novel deep learning model, called FCPNet, is proposed to achieve the fuel consumption prediction (FCP) before the flight departure. Considering the influential factors for aircraft performance, the multi-modal information sources, including the planned route, weather information, and operation details, are selected as the model input to predict fuel consumption. Correspondingly, three modules are innovatively proposed to learn embedding features from multi-modal inputs. Based on the planned route, the graph convolutional network is proposed to mine the spatial correlations in the non-Eulerian route network. Considering the grid attributes of the weather information, the ConvLSTM is applied to learn abstract representations from both the temporal and spatial dimensions, in which the three-dimensional convolution neural networks are also designed to fine-tune intermediate feature maps. The fully connected layer is also proposed to learn informative features from operation details. Finally, an attention-based fusion network is presented to generate the final embedding by considering the unique contributions of the multi-modality sources, which are further applied to predict flight fuel consumption. A binary encoding representation is proposed to formulate the FCP task as a multi-binary classification problem. The proposed model is validated on a real-world dataset, and the results demonstrate that it outperforms other baselines, i.e., achieving a 6.50% mean absolute percentage error, which can practically support the airline operation and global emission control before flight departure.

Occlusion facial expression recognition based on feature fusion residual attention network.

Recognizing occluded facial expressions in the wild poses a significant challenge. However, most previous approaches rely solely on either global or local feature-based methods, leading to the loss of relevant expression features. To address these issues, a feature fusion residual attention network (FFRA-Net) is proposed. FFRA-Net consists of a multi-scale module, a local attention module, and a feature fusion module. The multi-scale module divides the intermediate feature map into several sub-feature maps in an equal manner along the channel dimension. Then, a convolution operation is applied to each of these feature maps to obtain diverse global features. The local attention module divides the intermediate feature map into several sub-feature maps along the spatial dimension. Subsequently, a convolution operation is applied to each of these feature maps, resulting in the extraction of local key features through the attention mechanism. The feature fusion module plays a crucial role in integrating global and local expression features while also establishing residual links between inputs and outputs to compensate for the loss of fine-grained features. Last, two occlusion expression datasets (FM_RAF-DB and SG_RAF-DB) were constructed based on the RAF-DB dataset. Extensive experiments demonstrate that the proposed FFRA-Net achieves excellent results on four datasets: FM_RAF-DB, SG_RAF-DB, RAF-DB, and FERPLUS, with accuracies of 77.87%, 79.50%, 88.66%, and 88.97%, respectively. Thus, the approach presented in this paper demonstrates strong applicability in the context of occluded facial expression recognition (FER).

GABNet: global attention block for retinal OCT disease classification.

The retina represents a critical ocular structure. Of the various ophthalmic afflictions, retinal pathologies have garnered considerable scientific interest, owing to their elevated prevalence and propensity to induce blindness. Among clinical evaluation techniques employed in ophthalmology, optical coherence tomography (OCT) is the most commonly utilized, as it permits non-invasive, rapid acquisition of high-resolution, cross-sectional images of the retina. Timely detection and intervention can significantly abate the risk of blindness and effectively mitigate the national incidence rate of visual impairments. This study introduces a novel, efficient global attention block (GAB) for feed forward convolutional neural networks (CNNs). The GAB generates an attention map along three dimensions (height, width, and channel) for any intermediate feature map, which it then uses to compute adaptive feature weights by multiplying it with the input feature map. This GAB is a versatile module that can seamlessly integrate with any CNN, significantly improving its classification performance. Based on the GAB, we propose a lightweight classification network model, GABNet, which we develop on a UCSD general retinal OCT dataset comprising 108,312 OCT images from 4686 patients, including choroidal neovascularization (CNV), diabetic macular edema (DME), drusen, and normal cases. Notably, our approach improves the classification accuracy by 3.7% over the EfficientNetV2B3 network model. We further employ gradient-weighted class activation mapping (Grad-CAM) to highlight regions of interest on retinal OCT images for each class, enabling doctors to easily interpret model predictions and improve their efficiency in evaluating relevant models. With the increasing use and application of OCT technology in the clinical diagnosis of retinal images, our approach offers an additional diagnostic tool to enhance the diagnostic efficiency of clinical OCT retinal images.