Recognition Benchmark Research Articles

Robust transformer with locality inductive bias and feature normalization

Vision transformers have been demonstrated to yield state-of-the-art results on a variety of computer vision tasks using attention-based networks. However, research works in transformers mostly do not investigate robustness/accuracy trade-off, and they still struggle to handle adversarial perturbations. In this paper, we explore the robustness of vision transformers against adversarial perturbations and try to enhance their robustness/accuracy trade-off in white box attack settings. To this end, we propose Locality iN Locality (LNL) transformer model. We prove that the locality introduction to LNL contributes to the robustness performance since it aggregates local information such as lines, edges, shapes, and even objects. In addition, to further improve the robustness performance, we encourage LNL to extract training signal from the moments (a.k.a., mean and standard deviation) and the normalized features. We validate the effectiveness and generality of LNL by achieving state-of-the-art results in terms of accuracy and robustness metrics on German Traffic Sign Recognition Benchmark (GTSRB) and Canadian Institute for Advanced Research (CIFAR-10). More specifically, for traffic sign classification, the proposed LNL yields gains of 1.1% and 35% in terms of clean and robustness accuracy compared to the state-of-the-art studies.

A novel hybrid machine learning approach for traffic sign detection using CNN-GRNN

The traffic signal recognition model plays a significant role in the intelligent transportation model, as traffic signals aid the drivers to driving the more professional with awareness. The primary goal of this paper is to proposea model that works for the recognition and detection of traffic signals. This work proposes the pre-processing and segmentation approach applying machine learning techniques are occurred recent trends of study. Initially, the median filter & histogram equalization technique is utilized for pre-processing the traffic signal images, and also information of the figures being increased. The contrast of the figures upgraded, and information about the color shape of traffic signals are applied by the model. To localize the traffic signal in the obtained image, then this region of interest in traffic signal figures are extracted. The traffic signal recognition and classification experiments are managed depending on the German Traffic Signal Recognition Benchmark-(GTSRB). Various machine learning techniques such as Support Vector Machine (SVM), Extreme Learning Machine (ELM), Linear Discriminant Analysis (LDA), Principal Component Analysis (PCA), Convolutional neural network (CNN)- General Regression Neural Network (GRNN) is used for the classification process. Finally, the obtained results will be compare in terms of the performance metrics like accuracy, F1 score, kappa score, jaccard score, sensitivity, specificity, recall, and precision. The result shows that CNN-GRNN with ML techniques by attaining 99.41% accuracy compare to other intelligent methods. In this proposed technique is used for detecting and classifying various categories of traffic signals to improve the accuracy and effectiveness of the system.

Egypt Monuments Dataset version 1: A Scalable Benchmark for Image Classification and Monument Recognition

The success of machine learning (ML) as well as deep learning (DL) depends largely on data availability and quality. The system’s performance is frequently more affected by the amount and quality of its training data than by its architecture and training specifics. Consequently, demand exists for challenging datasets that both precisely measure performance and present unique challenges with real-world applications. The Egypt Monuments Dataset v1 (EGYPT-v1) is introduced as a new scalable benchmark for fine-image classification (IC) and object recognition (OR) in the domain of ancient Egyptian monuments. EGYPT-v1 dataset is by far the world’s first large specified such dataset to date, with over seven thousand images and 40 distinct instance labels. The dataset composes different categories of monuments such as pyramids, temples, mummies, statues, head statues, bust statues, heritage sites, palaces and shrines. Several advanced deep network architectures were tested to appraise the classification difficulty in the EGYPT-v1 dataset, namely ResNet50, Inception V3, and LeNet5 models. The models achieved accuracy rates as follows: 99.13%, 90.90%, and 92.64%, respectively. The dataset was predominantly created by manually collecting images from the popular global online video-sharing and social media platform, Youtube, as well as WATCHiT, Egypt’s top streaming entertainment service. Additionally, Wikimedia Commons, the largest crowdsourced media repository in the world, was used as a secondary source of images. The images that comprise the dataset can be accessed on the GitHub repository https://github.com/mennatallahhassan/egypt-monuments-dataset.

MAR20： A benchmark for military aircraft recognition in remote sensing images

MAR20： A benchmark for military aircraft recognition in remote sensing images

Transductive Prototypical Attention Reasoning Network for Few-Shot SAR Target Recognition

Deep learning-based synthetic aperture radar (SAR) automatic target recognition (ATR) algorithms have achieved outstanding performance under the condition of hundreds or thousands of training samples in recent years. Nevertheless, it is often rare to acquire great quantities of target samples in real SAR application scenarios. This article proposes a novel ATR method called transductive prototypical attention reasoning network (TPARN) to solve the problem of SAR target recognition with only a few training samples. To be specific, a region awareness-based feature extraction model is first developed, which can effectively focus on the target region of interest and suppress the background clutter by embedding direction-aware and position-sensitive information to extract more transferable knowledge. To heighten the discrimination of the sample features, a cross-feature spatial attention module is then proposed following the feature embedding model. Finally, a transductive prototype reasoning method is presented to realize the identity reasoning of the target, which can continuously update each class prototype with training samples and test samples together, thereby improving the classification accuracy. In addition, a marginal adaptive hybrid loss is proposed to obtain a discriminative feature embedding space with intra-class compactness and inter-class divergence, aiming to facilitate subsequent target identity reasoning. Extensive experiments on the moving and stationary target acquisition and recognition (MSTAR) benchmark dataset reveal that the proposed method outperforms some state-of-the-arts under different few-shot SAR ATR tasks.

Attribute-Guided Generative Adversarial Network With Improved Episode Training Strategy for Few-Shot SAR Image Generation

Deep learning-based models usually require a large amount of data for training, which guarantees the effectiveness of the trained model. Generative models are no exception, and sufficient training data are necessary for the diversity of generated images. However, for SAR images, data acquisition is expensive. Therefore, SAR image generation under few training samples is still a challenging problem to be solved. In this paper, we propose an attribute-guided generative adversarial network (AGGAN) with improved episode training strategy for few-shot SAR image generation. Firstly, we design the AGGAN structure, and spectral normalization is used to stabilize the training in the few-shot situation. The attribute labels of AGGAN are designed to be the category and aspect angle labels, which are essential information for SAR images. Secondly, an improved episode training strategy is proposed according to the characteristics of the few-shot generative task, and it can improve the quality of generated images in the few-shot situation. In addition, we explore the effectiveness of the proposed method when using different auxiliary data for training and use the Moving and Stationary Target Acquisition and Recognition (MSTAR) benchmark dataset and a simulated SAR dataset for verification. The experimental results show that AGGAN and the proposed improved episode training strategy can generate images of better quality when compared with some existing methods, which have been verified through visual observation, image similarity measures, and recognition experiments. When applying the generated images to the 5-shot SAR image recognition problem, the average recognition accuracy can be improved by at least 4 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> .

Contrastive Feature Disentangling for Partial Aspect Angles SAR Noncooperative Target Recognition

Deep learning algorithms have achieved state-of-the-art progress in synthetic aperture radar (SAR) automatic target recognition (ATR) tasks. They theoretically assume that training and test samples are independent and identically distributed (i.i.d.) for generalization, but it is intractable for practical ATR scenarios. In this paper, we propose a novel contrastive feature disentangling framework termed ConFeDent to learn features with improved generalization performance under a condition of a weaker distribution consistency. More specifically, ConFeDent aims to describe the semantic interactions between two arbitrary SAR training samples instead of treating them independently. It can implicitly disentangle features encoding the pose and identity knowledge from the whole samples with a semi-parametric geometric transformation model and a second-order energy model. In particular, except for the identity label, we use deductive-based geometry knowledge as supervision to teach the model to learn the concept of aspect angle variation. A progressively amortized inference scheme is constructed for efficient feature learning and recognition in an end-to-end manner. Finally, we further release a strengthened version, called ConFeDent+, which can explicitly utilize and learn more information from cross-category samples. Experimental results on the moving and stationary target acquisition and recognition (MSTAR) benchmark demonstrate the effectiveness of our proposed models in the SAR ATR. In particular, we validate the algorithms in a more challenging scenario where the range of aspect angles for training and testing samples is permitted to be disparate. Our model can achieve much higher recognition accuracy than other SAR ATR algorithms.

Joint Holistic and Masked Face Recognition

With the widespread use of face masks due to the COVID-19 pandemic, accurate masked face recognition has become more crucial than ever. While several studies have investigated masked face recognition using convolutional neural networks (CNNs), there is a paucity of research exploring the use of plain Vision Transformers (ViTs) for this task. Unlike ViT models used in image classification, object detection, and semantic segmentation, the model trained by modern face recognition losses struggles to converge when trained from scratch. To this end, this paper initializes the model parameters via a proxy task of patch reconstruction and observes that the ViT backbone exhibits improved training stability with satisfactory performance for face recognition. Beyond the training stability, two strategies based on prompts are proposed to integrate holistic and masked face recognition in a single framework, namely FaceT. Along with popular holistic face recognition benchmarks, several open-sourced masked face recognition benchmarks are collected for evaluation. Our extensive experiments demonstrate that the proposed FaceT performs on par or better than state-of-the-art CNNs on both holistic and masked face recognition benchmarks. Codes will be made available at https://github.com/zyainfal/Joint-Holistic-and-Masked-Face-Recognition.

Knowledge-aware Global Reasoning for Situation Recognition.

The task of situation recognition aims to solve the visual reasoning problem with the ability to predict the activity happening (salient action) in an image and the nouns of all associated semantic roles playing in the activity. This poses severe challenges due to long-tailed data distributions and local class ambiguities. Prior works only propagate the local noun-level features on one single image without utilizing global information. We propose a Knowledge-aware Global Reasoning (KGR) framework to endow neural networks with the capability of adaptive global reasoning over nouns by exploiting diverse statistical knowledge. Our KGR is a local-global architecture, which consists of a local encoder to generate noun features using local relations and a global encoder to enhance the noun features via global reasoning supervised by an external global knowledge pool. The global knowledge pool is created by counting the pairwise relationships of nouns in the dataset. In this paper, we design an action-guided pairwise knowledge as the global knowledge pool based on the characteristic of the situation recognition task. Extensive experiments have shown that our KGR not only achieves state-of-the-art results on a large-scale situation recognition benchmark, but also effectively solves the long-tailed problem of noun classification by our global knowledge. Code for KGR is available: https://github.com/haofanwang/KGRN-SR.

Collaborative Multilingual Continuous Sign Language Recognition: A Unified Framework

Current continuous sign language recognition systems generally target on a single language. When it comes to the multilingual problem, existing solutions often build separate models based on the same network and then train them with their corresponding sign language corpora. Observing that different sign languages share some low-level visual patterns, we argue that it is beneficial to optimize the recognition model in a collaborative way. With this motivation, we propose the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">first</i> unified framework for multilingual continuous sign language recognition. Our framework consists of a shared visual encoder for visual information encoding, multiple language-dependent sequential modules for long-range temporal dependency learning aimed at different languages, and a universal sequential module to learn the commonality of all languages. An additional language embedding is introduced to distinguish different languages within the shared temporal encoders. Further, we present a max-probability decoding method to obtain the alignment between sign videos and sign words for visual encoder refinement. We evaluate our approach on three continuous sign language recognition benchmarks, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e.,</i> RWTH-PHOENIX-Weather, CSL and GSL-SD. The experimental results reveal that our method outperforms the individually trained recognition models. Our method also demonstrates better performance compared with state-of-the-art algorithms.

Wave interference network with a wave function for traffic sign recognition.

In this paper, we successfully combine convolution with a wave function to build an effective and efficient classifier for traffic signs, named the wave interference network (WiNet). In the WiNet, the feature map extracted by the convolutional filters is refined into many entities from an input image. Each entity is represented as a wave. We utilize Euler's formula to unfold the wave function. Based on the wave-like information representation, the model modulates the relationship between the entities and the fixed weights of convolution adaptively. Experiment results on the Chinese Traffic Sign Recognition Database (CTSRD) and the German Traffic Sign Recognition Benchmark (GTSRB) demonstrate that the performance of the presented model is better than some other models, such as ResMLP, ResNet50, PVT and ViT in the following aspects: 1) WiNet obtains the best accuracy rate with 99.80% on the CTSRD and recognizes all images exactly on the GTSRB; 2) WiNet gains better robustness on the dataset with different noises compared with other models; 3) WiNet has a good generalization on different datasets.

Robust tactile object recognition in open-set scenarios using Gaussian prototype learning.

Tactile object recognition is crucial for effective grasping and manipulation. Recently, it has started to attract increasing attention in robotic applications. While there are many works on tactile object recognition and they also achieved promising performances in some applications, most of them are usually limited to closed world scenarios, where the object instances to be recognition in deployment are known and the same as that of during training. Since robots usually operate in realistic open-set scenarios, they inevitably encounter unknown objects. If automation systems falsely recognize unknown objects as one of the known classes based on the pre-trained model, it can lead to potentially catastrophic consequences. It motivates us to break the closed world assumption and to study tactile object recognition in realistic open-set conditions. Although several open-set recognition methods have been proposed, they focused on visual tasks and may not be suitable for tactile recognition. It is mainly due to that these methods do not take into account the special characteristic of tactile data in their models. To this end, we develop a novel Gaussian Prototype Learning method for robust tactile object recognition. Particularly, the proposed method converts feature distributions to probabilistic representations, and exploit uncertainty for tactile recognition in open-set scenarios. Experiments on the two tactile recognition benchmarks demonstrate the effectiveness of the proposed method on open-set tasks.

Efficient Transformer-Based Compressed Video Modeling via Informative Patch Selection.

Recently, Transformer-based video recognition models have achieved state-of-the-art results on major video recognition benchmarks. However, their high inference cost significantly limits research speed and practical use. In video compression, methods considering small motions and residuals that are less informative and assigning short code lengths to them (e.g., MPEG4) have successfully reduced the redundancy of videos. Inspired by this idea, we propose Informative Patch Selection (IPS), which efficiently reduces the inference cost by excluding redundant patches from the input of the Transformer-based video model. The redundancy of each patch is calculated from motions and residuals obtained while decoding a compressed video. The proposed method is simple and effective in that it can dynamically reduce the inference cost depending on the input without any policy model or additional loss term. Extensive experiments on action recognition demonstrated that our method could significantly improve the trade-off between the accuracy and inference cost of the Transformer-based video model. Although the method does not require any policy model or additional loss term, its performance approaches that of existing methods that do require them.

SeVuc: A study on the Security Vulnerabilities of Capsule Networks against adversarial attacks

Capsule Networks (CapsNets) preserve the hierarchical spatial relationships between objects, and thereby bear the potential to surpass the performance of traditional Convolutional Neural Networks (CNNs) in performing tasks like image classification. This makes CapsNets suitable for the smart cyber–physical systems (CPS), where a large amount of training data may not be available. A large body of work has explored adversarial examples for CNNs, but their effectiveness on CapsNets has not yet been studied systematically. In our work, we perform an analysis to study the vulnerabilities in CapsNets to adversarial attacks. These perturbations, added to the test inputs, are small and imperceptible to humans, but can fool the network to mispredict. We propose a greedy algorithm to automatically generate imperceptible adversarial examples in a black-box attack scenario. We show that this kind of attacks, when applied to the German Traffic Sign Recognition Benchmark and CIFAR10 datasets, mislead CapsNets in making a correct classification, which can be catastrophic for smart CPS, like autonomous vehicles. Moreover, we apply the same kind of adversarial attacks to a 5-layer CNN (LeNet), to a 9-layer CNN (VGGNet), and to a 20-layer CNN (ResNet), and analyze the outcome, compared to the CapsNets, to study their different behaviors under the adversarial attacks.

Lightweight Semantic-Guided Neural Networks Based on Single Head Attention for Action Recognition

Skeleton-based action recognition can achieve a relatively high performance by transforming the human skeleton structure in an image into a graph and applying action recognition based on structural changes in the body. Among the many graph convolutional network (GCN) approaches used in skeleton-based action recognition, semantic-guided neural networks (SGNs) are fast action recognition algorithms that hierarchically learn spatial and temporal features by applying a GCN. However, because an SGN focuses on global feature learning rather than local feature learning owing to the structural characteristics, there is a limit to an action recognition in which the dependency between neighbouring nodes is important. To solve these problems and simultaneously achieve a real-time action recognition in low-end devices, in this study, a single head attention (SHA) that can overcome the limitations of an SGN is proposed, and a new SGN-SHA model that combines SHA with an SGN is presented. In experiments on various action recognition benchmark datasets, the proposed SGN-SHA model significantly reduced the computational complexity while exhibiting a performance similar to that of an existing SGN and other state-of-the-art methods.

Classification by Principal Component Regression in the Real and Hypercomplex Domains

Linear regression is a simple and widely used machine learning algorithm. It is a statistical approach for modeling the relationship between a scalar variable and one or more variables. In this paper, a classification by principal component regression (CbPCR) strategy is proposed. This strategy depends on performing regression of each data class in terms of its principal components. This CbPCR formulation leads to a new formulation of the Linear Regression Classification (LRC) problem that preserves the key information of the data classes while providing more compact closed-form solutions. For the sake of image classification, this strategy is also extended to the 4D hypercomplex domains to take into account the color information of the image. Quaternion and reduced biquaternion CbPCR strategies are proposed by representing each channel of the color image as one of the imaginary parts of a quaternion or reduced biquaternion number. Experiments on two color face recognition benchmark databases show that the proposed methods achieve better accuracies by a margin of about 3% over the original LRC and like methods.

Traffic sign classification using CNN and detection using faster-RCNN and YOLOV4

Autonomous driving cars are becoming popular everywhere and the need for a robust traffic sign recognition system that ensures safety by recognizing traffic signs accurately and fast is increasing. In this paper, we build a CNN that can classify 43 different traffic signs from the German Traffic Sign Recognition benchmark dataset. The dataset is made up of 39,186 images for training and 12,630 for testing. Our CNN for classification is light and reached an accuracy of 99.20% with only 0.8 M parameters. It is tested also under severe conditions to prove its generalization ability. We also used Faster R–CNN and YOLOv4 networks to implement a recognition system for traffic signs. The German Traffic Sign Detection benchmark dataset was used. Faster R–CNN obtained a mean average precision (mAP) of 43.26% at 6 Frames Per Second (FPS), which is not suitable for real-time application. YOLOv4 achieved an mAP of 59.88% at 35 FPS, which is the preferred model for real-time traffic sign detection. These mAPs are obtained using Intersect Over Union of 50%. A comparative analysis is also presented between these models.

Multi-Task Learning for Scene Text Image Super-Resolution with Multiple Transformers

Scene text image super-resolution aims to improve readability by recovering text shapes from low-resolution degraded text images. Although recent developments in deep learning have greatly improved super-resolution (SR) techniques, recovering text images with irregular shapes, heavy noise, and blurriness is still challenging. This is because networks with Convolutional Neural Network (CNN)-based backbones cannot sufficiently capture the global long-range correlations of text images or detailed sequential information about the text structure. In order to address this issue, this paper proposes a Multi-task learning-based Text Super-resolution (MTSR) Network to approach this problem. MTSR is a multi-task architecture for image reconstruction and SR. It uses transformer-based modules to transfer complementary features of the reconstruction model, such as noise removal capability and text structure information, to the SR model. In addition, another transformer-based module using 2D positional encoding is used to handle irregular deformations of the text. The feature maps generated from these two transformer-based modules are fused to attempt improvement of the visual quality of images with heavy noise, blurriness, and irregular deformations. Experimental results on the TextZoom dataset and several scene text recognition benchmarks show that our MTSR significantly improves the accuracy of existing text recognizers.

DANet: Semi-supervised differentiated auxiliaries guided network for video action recognition

Video Action Recognition (ViAR) aims to identify the category of the human action observed in a given video. With the advent of Deep Learning (DL) techniques, noticeable performance breakthroughs have been achieved in this study. However, the success of most existing DL-based ViAR methods heavily relies on the existence of a large amount of annotated data, i.e., videos with corresponding action categories. In practice, obtaining such a desired number of annotations is often difficult due to expensive labeling costs, which may lead to significant performance degradation for these methods. To address this issue, we propose an end-to-end semi-supervised Differentiated Auxiliary guided Network (DANet) to best use a few annotated videos. Except for the common supervised learning on a few annotated videos, the DANet also involves the knowledge of multiple pre-trained auxiliary networks to optimize the ViAR network in a self-supervised way on the unannotated data by removing the annotations. Considering the tight connection between video action recognition and classical static image-based visual tasks, the abundant knowledge from the pre-trained static image-based models can be used for training the ViAR model. Specifically, the DANet is a two-branch architecture, which includes a target branch of the ViAR network, and an auxiliary branch of multiple auxiliary networks (i.e., referring to diverse off-the-shelf models of relevant image tasks). Given a limited number of annotated videos, we train the target ViAR network end-to-end in a semi-supervised way, namely, with both the supervised cross-entropy loss on annotated videos, and the per-auxiliary weighted self-supervised contrastive losses on the same videos but without using annotations. Besides, we further explore different weighted guidance of the auxiliary networks to the ViAR network to better reflect different relationships between the image-based models and the ViAR model. Finally, we conduct extensive experiments on several popular action recognition benchmarks in comparison with existing state-of-the-art methods, and the experimental results demonstrate the superiority of DANet over most of the compared methods. In particular, the DANet obviously suppresses state-of-the-art ViAR methods even with very fewer annotated videos.

Braille letter reading: A benchmark for spatio-temporal pattern recognition on neuromorphic hardware.

Spatio-temporal pattern recognition is a fundamental ability of the brain which is required for numerous real-world activities. Recent deep learning approaches have reached outstanding accuracies in such tasks, but their implementation on conventional embedded solutions is still very computationally and energy expensive. Tactile sensing in robotic applications is a representative example where real-time processing and energy efficiency are required. Following a brain-inspired computing approach, we propose a new benchmark for spatio-temporal tactile pattern recognition at the edge through Braille letter reading. We recorded a new Braille letters dataset based on the capacitive tactile sensors of the iCub robot's fingertip. We then investigated the importance of spatial and temporal information as well as the impact of event-based encoding on spike-based computation. Afterward, we trained and compared feedforward and recurrent Spiking Neural Networks (SNNs) offline using Backpropagation Through Time (BPTT) with surrogate gradients, then we deployed them on the Intel Loihi neuromorphic chip for fast and efficient inference. We compared our approach to standard classifiers, in particular to the Long Short-Term Memory (LSTM) deployed on the embedded NVIDIA Jetson GPU, in terms of classification accuracy, power, and energy consumption together with computational delay. Our results show that the LSTM reaches ~97% of accuracy, outperforming the recurrent SNN by ~17% when using continuous frame-based data instead of event-based inputs. However, the recurrent SNN on Loihi with event-based inputs is ~500 times more energy-efficient than the LSTM on Jetson, requiring a total power of only ~30 mW. This work proposes a new benchmark for tactile sensing and highlights the challenges and opportunities of event-based encoding, neuromorphic hardware, and spike-based computing for spatio-temporal pattern recognition at the edge.