Vision Transformers (ViTs) have emerged as popular models in computer vision, demonstrating state-of-the-art performance across various tasks. This success typically follows a two-stage strategy involving pre-training on large-scale datasets using self-supervised signals, such as masked random patches, followed by fine-tuning on task-specific labeled datasets with cross-entropy loss. However, this reliance on cross-entropy loss has been identified as a limiting factor in ViTs, affecting their generalization and transferability to downstream tasks. Addressing this critical challenge, we introduce a novel Label-aware Contrastive Training framework, LaCViT, which significantly enhances the quality of embeddings in ViTs. LaCViT not only addresses the limitations of cross-entropy loss but also facilitates more effective transfer learning across diverse image classification tasks. Our comprehensive experiments on eight standard image classification datasets reveal that LaCViT statistically significantly enhances the performance of three evaluated ViTs by up-to 10.78% under Top-1 Accuracy.
When dealing with the task of fine-grained scene image classification, most previous works lay much emphasis on global visual features when doing multi-modal feature fusion. In other words, models are deliberately designed based on prior intuitions about the importance of different modalities. In this paper, we present a new multi-modal feature fusion approach named MAA (Modality-Agnostic Adapter), trying to make the model learn the importance of different modalities in different cases adaptively, without giving a prior setting in the model architecture. More specifically, we eliminate the modal differences in distribution and then use a modality-agnostic Transformer encoder for a semantic-level feature fusion. Our experiments demonstrate that MAA achieves state-of-the-art results on benchmarks by applying the same modalities with previous methods. Besides, it is worth mentioning that new modalities can be easily added when using MAA and further boost the performance. Code is available at https://github.com/quniLcs/MAA.
Coronavirus disease 2019 (COVID-19) is one of the most destructive pandemic after millennium, forcing the world to tackle a health crisis. Automated lung infections classification using chest X-ray (CXR) images could strengthen diagnostic capability when handling COVID-19. However, classifying COVID-19 from pneumonia cases using CXR image is a difficult task because of shared spatial characteristics, high feature variation and contrast diversity between cases. Moreover, massive data collection is impractical for a newly emerged disease, which limited the performance of data thirsty deep learning models. To address these challenges, Multiscale Attention Guided deep network with Soft Distance regularization (MAG-SD) is proposed to automatically classify COVID-19 from pneumonia CXR images. In MAG-SD, MA-Net is used to produce prediction vector and attention from multiscale feature maps. To improve the robustness of trained model and relieve the shortage of training data, attention guided augmentations along with a soft distance regularization are posed, which aims at generating meaningful augmentations and reduce noise. Our multiscale attention model achieves better classification performance on our pneumonia CXR image dataset. Plentiful experiments are proposed for MAG-SD which demonstrates its unique advantage in pneumonia classification over cutting-edge models. The code is available at https://github.com/JasonLeeGHub/MAG-SD.
Recent researches on unsupervised person re-identification~(reID) have demonstrated that pre-training on unlabeled person images achieves superior performance on downstream reID tasks than pre-training on ImageNet. However, those pre-trained methods are specifically designed for reID and suffer flexible adaption to other pedestrian analysis tasks. In this paper, we propose VAL-PAT, a novel framework that learns transferable representations to enhance various pedestrian analysis tasks with multimodal information. To train our framework, we introduce three learning objectives, \emph{i.e.,} self-supervised contrastive learning, image-text contrastive learning and multi-attribute classification. The self-supervised contrastive learning facilitates the learning of the intrinsic pedestrian properties, while the image-text contrastive learning guides the model to focus on the appearance information of pedestrians.Meanwhile, multi-attribute classification encourages the model to recognize attributes to excavate fine-grained pedestrian information. We first perform pre-training on LUPerson-TA dataset, where each image contains text and attribute annotations, and then transfer the learned representations to various downstream tasks, including person reID, person attribute recognition and text-based person search. Extensive experiments demonstrate that our framework facilitates the learning of general pedestrian representations and thus leads to promising results on various pedestrian analysis tasks.
Few-shot classification aims to recognize novel categories with only few labeled images in each class. Existing metric-based few-shot classification algorithms predict categories by comparing the feature embeddings of query images with those from a few labeled images (support examples) using a learned metric function. While promising performance has been demonstrated, these methods often fail to generalize to unseen domains due to large discrepancy of the feature distribution across domains. In this work, we address the problem of few-shot classification under domain shifts for metric-based methods. Our core idea is to use feature-wise transformation layers for augmenting the image features using affine transforms to simulate various feature distributions under different domains in the training stage. To capture variations of the feature distributions under different domains, we further apply a learning-to-learn approach to search for the hyper-parameters of the feature-wise transformation layers. We conduct extensive experiments and ablation studies under the domain generalization setting using five few-shot classification datasets: mini-ImageNet, CUB, Cars, Places, and Plantae. Experimental results demonstrate that the proposed feature-wise transformation layer is applicable to various metric-based models, and provides consistent improvements on the few-shot classification performance under domain shift.
Vision mambas have demonstrated strong performance with linear complexity to the number of vision tokens. Their efficiency results from processing image tokens sequentially. However, most existing methods employ patch-based image tokenization and then flatten them into 1D sequences for causal processing, which ignore the intrinsic 2D structural correlations of images. It is also difficult to extract global information by sequential processing of local patches. In this paper, we propose a global image serialization method to transform the image into a sequence of causal tokens, which contain global information of the 2D image. We first convert the image from the spatial domain to the frequency domain using Discrete Cosine Transform (DCT) and then arrange the pixels with corresponding frequency ranges. We further transform each set within the same frequency band back to the spatial domain to obtain a series of images before tokenization. We construct a vision mamba model, GlobalMamba, with a causal input format based on the proposed global image serialization, which can better exploit the causal relations among image sequences. Extensive experiments demonstrate the effectiveness of our GlobalMamba, including image classification on ImageNet-1K, object detection on COCO, and semantic segmentation on ADE20K.
A novel approach of data augmentation based on irregular superpixel decomposition is proposed. This approach called SuperpixelGridMasks permits to extend original image datasets that are required by training stages of machine learning-related analysis architectures towards increasing their performances. Three variants named SuperpixelGridCut, SuperpixelGridMean and SuperpixelGridMix are presented. These grid-based methods produce a new style of image transformations using the dropping and fusing of information. Extensive experiments using various image classification models and datasets show that baseline performances can be significantly outperformed using our methods. The comparative study also shows that our methods can overpass the performances of other data augmentations. Experimental results obtained over image recognition datasets of varied natures show the efficiency of these new methods. SuperpixelGridCut, SuperpixelGridMean and SuperpixelGridMix codes are publicly available at https://github.com/hammoudiproject/SuperpixelGridMasks
ResNets (or Residual Networks) are one of the most commonly used models for image classification tasks. In this project, we design and train a modified ResNet model for CIFAR-10 image classification. In particular, we aimed at maximizing the test accuracy on the CIFAR-10 benchmark while keeping the size of our ResNet model under the specified fixed budget of 5 million trainable parameters. Model size, typically measured as the number of trainable parameters, is important when models need to be stored on devices with limited storage capacity (e.g. IoT/edge devices). In this article, we present our residual network design which has less than 5 million parameters. We show that our ResNet achieves a test accuracy of 96.04% on CIFAR-10 which is much higher than ResNet18 (which has greater than 11 million trainable parameters) when equipped with a number of training strategies and suitable ResNet hyperparameters. Models and code are available at https://github.com/Nikunj-Gupta/Efficient_ResNets.
We apply pre-trained Vision Transformers (ViTs), originally developed for image recognition, to the analysis of astronomical spectral data. By converting traditional one-dimensional spectra into two-dimensional image representations, we enable ViTs to capture both local and global spectral features through spatial self-attention. We fine-tune a ViT pretrained on ImageNet using millions of spectra from the SDSS and LAMOST surveys, represented as spectral plots. Our model is evaluated on key tasks including stellar object classification and redshift ($z$) estimation, where it demonstrates strong performance and scalability. We achieve classification accuracy higher than Support Vector Machines and Random Forests, and attain $R^2$ values comparable to AstroCLIP's spectrum encoder, even when generalizing across diverse object types. These results demonstrate the effectiveness of using pretrained vision models for spectroscopic data analysis. To our knowledge, this is the first application of ViTs to large-scale, which also leverages real spectroscopic data and does not rely on synthetic inputs.
We introduce a novel method to combat label noise when training deep neural networks for classification. We propose a loss function that permits abstention during training thereby allowing the DNN to abstain on confusing samples while continuing to learn and improve classification performance on the non-abstained samples. We show how such a deep abstaining classifier (DAC) can be used for robust learning in the presence of different types of label noise. In the case of structured or systematic label noise -- where noisy training labels or confusing examples are correlated with underlying features of the data-- training with abstention enables representation learning for features that are associated with unreliable labels. In the case of unstructured (arbitrary) label noise, abstention during training enables the DAC to be used as an effective data cleaner by identifying samples that are likely to have label noise. We provide analytical results on the loss function behavior that enable dynamic adaption of abstention rates based on learning progress during training. We demonstrate the utility of the deep abstaining classifier for various image classification tasks under different types of label noise; in the case of arbitrary label noise, we show significant improvements over previously published results on multiple image benchmarks. Source code is available at https://github.com/thulas/dac-label-noise
Despite the potential of synthetic medical data for augmenting and improving the generalizability of deep learning models, memorization in generative models can lead to unintended leakage of sensitive patient information and limit model utility. Thus, the use of memorizing generative models in the medical domain can jeopardize patient privacy. We propose a framework for identifying replicas, i.e. nearly identical copies of the training data, in synthetic medical image datasets. Our REpLIca deteCTion (RELICT) framework for medical image generative models evaluates image similarity using three complementary approaches: (1) voxel-level analysis, (2) feature-level analysis by a pretrained medical foundation model, and (3) segmentation-level analysis. Two clinically relevant 3D generative modelling use cases were investigated: non-contrast head CT with intracerebral hemorrhage (N=774) and time-of-flight MR angiography of the Circle of Willis (N=1,782). Expert visual scoring was used as the reference standard to assess the presence of replicas. We report the balanced accuracy at the optimal threshold to assess replica classification performance. The reference visual rating identified 45 of 50 and 5 of 50 generated images as replicas for the NCCT and TOF-MRA use cases, respectively. Image-level and feature-level measures perfectly classified replicas with a balanced accuracy of 1 when an optimal threshold was selected for the NCCT use case. A perfect classification of replicas for the TOF-MRA case was not possible at any threshold, with the segmentation-level analysis achieving a balanced accuracy of 0.79. Replica detection is a crucial but neglected validation step for the development of generative models in medical imaging. The proposed RELICT framework provides a standardized, easy-to-use tool for replica detection and aims to facilitate responsible and ethical medical image synthesis.
We propose a novel loss function for imbalanced classification. LDAM loss, which minimizes a margin-based generalization bound, is widely utilized for class-imbalanced image classification. Although, by using LDAM loss, it is possible to obtain large margins for the minority classes and small margins for the majority classes, the relevance to a large margin, which is included in the original softmax cross entropy loss, is not be clarified yet. In this study, we reconvert the formula of LDAM loss using the concept of the large margin softmax cross entropy loss based on the softplus function and confirm that LDAM loss includes a wider large margin than softmax cross entropy loss. Furthermore, we propose a novel Enlarged Large Margin (ELM) loss, which can further widen the large margin of LDAM loss. ELM loss utilizes the large margin for the maximum logit of the incorrect class in addition to the basic margin used in LDAM loss. Through experiments conducted on imbalanced CIFAR datasets and large-scale datasets with long-tailed distribution, we confirmed that classification accuracy was much improved compared with LDAM loss and conventional losses for imbalanced classification.
For safety-related applications, it is crucial to produce trustworthy deep neural networks whose prediction is associated with confidence that can represent the likelihood of correctness for subsequent decision-making. Existing dense binary classification models are prone to being over-confident. To improve model calibration, we propose Adaptive Stochastic Label Perturbation (ASLP) which learns a unique label perturbation level for each training image. ASLP employs our proposed Self-Calibrating Binary Cross Entropy (SC-BCE) loss, which unifies label perturbation processes including stochastic approaches (like DisturbLabel), and label smoothing, to correct calibration while maintaining classification rates. ASLP follows Maximum Entropy Inference of classic statistical mechanics to maximise prediction entropy with respect to missing information. It performs this while: (1) preserving classification accuracy on known data as a conservative solution, or (2) specifically improves model calibration degree by minimising the gap between the prediction accuracy and expected confidence of the target training label. Extensive results demonstrate that ASLP can significantly improve calibration degrees of dense binary classification models on both in-distribution and out-of-distribution data. The code is available on https://github.com/Carlisle-Liu/ASLP.
In this paper, we present Co-scale conv-attentional image Transformers (CoaT), a Transformer-based image classifier equipped with co-scale and conv-attentional mechanisms. First, the co-scale mechanism maintains the integrity of Transformers' encoder branches at individual scales, while allowing representations learned at different scales to effectively communicate with each other; we design a series of serial and parallel blocks to realize the co-scale mechanism. Second, we devise a conv-attentional mechanism by realizing a relative position embedding formulation in the factorized attention module with an efficient convolution-like implementation. CoaT empowers image Transformers with enriched multi-scale and contextual modeling capabilities. On ImageNet, relatively small CoaT models attain superior classification results compared with similar-sized convolutional neural networks and image/vision Transformers. The effectiveness of CoaT's backbone is also illustrated on object detection and instance segmentation, demonstrating its applicability to downstream computer vision tasks.
To better understand scene images in the field of remote sensing, multi-label annotation of scene images is necessary. Moreover, to enhance the performance of deep learning models for dealing with semantic scene understanding tasks, it is vital to train them on large-scale annotated data. However, most existing datasets are annotated by a single label, which cannot describe the complex remote sensing images well because scene images might have multiple land cover classes. Few multi-label high spatial resolution remote sensing datasets have been developed to train deep learning models for multi-label based tasks, such as scene classification and image retrieval. To address this issue, in this paper, we construct a multi-label high spatial resolution remote sensing dataset named MLRSNet for semantic scene understanding with deep learning from the overhead perspective. It is composed of high-resolution optical satellite or aerial images. MLRSNet contains a total of 109,161 samples within 46 scene categories, and each image has at least one of 60 predefined labels. We have designed visual recognition tasks, including multi-label based image classification and image retrieval, in which a wide variety of deep learning approaches are evaluated with MLRSNet. The experimental results demonstrate that MLRSNet is a significant benchmark for future research, and it complements the current widely used datasets such as ImageNet, which fills gaps in multi-label image research. Furthermore, we will continue to expand the MLRSNet. MLRSNet and all related materials have been made publicly available at https://data.mendeley.com/datasets/7j9bv9vwsx/2 and https://github.com/cugbrs/MLRSNet.git.
Machine learning technologies using deep neural networks (DNNs), especially convolutional neural networks (CNNs), have made automated, accurate, and fast medical image analysis a reality for many applications, and some DNN-based medical image analysis systems have even been FDA-cleared. Despite the progress, challenges remain to build DNNs as reliable as human expert doctors. It is known that DNN classifiers may not be robust to noises: by adding a small amount of noise to an input image, a DNN classifier may make a wrong classification of the noisy image (i.e., in-distribution adversarial sample), whereas it makes the right classification of the clean image. Another issue is caused by out-of-distribution samples that are not similar to any sample in the training set. Given such a sample as input, the output of a DNN will become meaningless. In this study, we investigated the in-distribution (IND) and out-of-distribution (OOD) adversarial robustness of a representative CNN for lumbar disk shape reconstruction from spine MR images. To study the relationship between dataset size and robustness to IND adversarial attacks, we used a data augmentation method to create training sets with different levels of shape variations. We utilized the PGD-based algorithm for IND adversarial attacks and extended it for OOD adversarial attacks to generate OOD adversarial samples for model testing. The results show that IND adversarial training can improve the CNN robustness to IND adversarial attacks, and larger training datasets may lead to higher IND robustness. However, it is still a challenge to defend against OOD adversarial attacks.
Masked image modeling (MIM) learns representations with remarkably good fine-tuning performances, overshadowing previous prevalent pre-training approaches such as image classification, instance contrastive learning, and image-text alignment. In this paper, we show that the inferior fine-tuning performance of these pre-training approaches can be significantly improved by a simple post-processing in the form of feature distillation (FD). The feature distillation converts the old representations to new representations that have a few desirable properties just like those representations produced by MIM. These properties, which we aggregately refer to as optimization friendliness, are identified and analyzed by a set of attention- and optimization-related diagnosis tools. With these properties, the new representations show strong fine-tuning performance. Specifically, the contrastive self-supervised learning methods are made as competitive in fine-tuning as the state-of-the-art masked image modeling (MIM) algorithms. The CLIP models' fine-tuning performance is also significantly improved, with a CLIP ViT-L model reaching 89.0% top-1 accuracy on ImageNet-1K classification. On the 3-billion-parameter SwinV2-G model, the fine-tuning accuracy is improved by +1.5 mIoU / +1.1 mAP to 61.4 mIoU / 64.2 mAP on ADE20K semantic segmentation and COCO object detection, respectively, creating new records on both benchmarks. More importantly, our work provides a way for the future research to focus more effort on the generality and scalability of the learnt representations without being pre-occupied with optimization friendliness since it can be enhanced rather easily. The code will be available at https://github.com/SwinTransformer/Feature-Distillation.
Deep learning motivated by convolutional neural networks has been highly successful in a range of medical imaging problems like image classification, image segmentation, image synthesis etc. However for validation and interpretability, not only do we need the predictions made by the model but also how confident it is while making those predictions. This is important in safety critical applications for the people to accept it. In this work, we used an encoder decoder architecture based on variational inference techniques for segmenting brain tumour images. We evaluate our work on the publicly available BRATS dataset using Dice Similarity Coefficient (DSC) and Intersection Over Union (IOU) as the evaluation metrics. Our model is able to segment brain tumours while taking into account both aleatoric uncertainty and epistemic uncertainty in a principled bayesian manner.
Optical satellite sensors cannot see the Earth's surface through clouds. Despite the periodic revisit cycle, image sequences acquired by Earth observation satellites are therefore irregularly sampled in time. State-of-the-art methods for crop classification (and other time series analysis tasks) rely on techniques that implicitly assume regular temporal spacing between observations, such as recurrent neural networks (RNNs). We propose to use neural ordinary differential equations (NODEs) in combination with RNNs to classify crop types in irregularly spaced image sequences. The resulting ODE-RNN models consist of two steps: an update step, where a recurrent unit assimilates new input data into the model's hidden state; and a prediction step, in which NODE propagates the hidden state until the next observation arrives. The prediction step is based on a continuous representation of the latent dynamics, which has several advantages. At the conceptual level, it is a more natural way to describe the mechanisms that govern the phenological cycle. From a practical point of view, it makes it possible to sample the system state at arbitrary points in time, such that one can integrate observations whenever they are available, and extrapolate beyond the last observation. Our experiments show that ODE-RNN indeed improves classification accuracy over common baselines such as LSTM, GRU, and temporal convolution. The gains are most prominent in the challenging scenario where only few observations are available (i.e., frequent cloud cover). Moreover, we show that the ability to extrapolate translates to better classification performance early in the season, which is important for forecasting.
In medicine, curated image datasets often employ discrete labels to describe what is known to be a continuous spectrum of healthy to pathological conditions, such as e.g. the Alzheimer's Disease Continuum or other areas where the image plays a pivotal point in diagnosis. We propose an architecture for image stratification based on a conditional variational autoencoder. Our framework, VAESim, leverages a continuous latent space to represent the continuum of disorders and finds clusters during training, which can then be used for image/patient stratification. The core of the method learns a set of prototypical vectors, each associated with a cluster. First, we perform a soft assignment of each data sample to the clusters. Then, we reconstruct the sample based on a similarity measure between the sample embedding and the prototypical vectors of the clusters. To update the prototypical embeddings, we use an exponential moving average of the most similar representations between actual prototypes and samples in the batch size. We test our approach on the MNIST-handwritten digit dataset and on a medical benchmark dataset called PneumoniaMNIST. We demonstrate that our method outperforms baselines in terms of kNN accuracy measured on a classification task against a standard VAE (up to 15% improvement in performance) in both datasets, and also performs at par with classification models trained in a fully supervised way. We also demonstrate how our model outperforms current, end-to-end models for unsupervised stratification.