Medical image classification has developed rapidly under the impetus of the convolutional neural network (CNN). Due to the fixed size of the receptive field of the convolution kernel, it is difficult to capture the global features of medical images. Although the self-attention-based Transformer can model long-range dependencies, it has high computational complexity and lacks local inductive bias. Much research has demonstrated that global and local features are crucial for image classification. However, medical images have a lot of noisy, scattered features, intra-class variation, and inter-class similarities. This paper proposes a three-branch hierarchical multi-scale feature fusion network structure termed as HiFuse for medical image classification as a new method. It can fuse the advantages of Transformer and CNN from multi-scale hierarchies without destroying the respective modeling so as to improve the classification accuracy of various medical images. A parallel hierarchy of local and global feature blocks is designed to efficiently extract local features and global representations at various semantic scales, with the flexibility to model at different scales and linear computational complexity relevant to image size. Moreover, an adaptive hierarchical feature fusion block (HFF block) is designed to utilize the features obtained at different hierarchical levels comprehensively. The HFF block contains spatial attention, channel attention, residual inverted MLP, and shortcut to adaptively fuse semantic information between various scale features of each branch. The accuracy of our proposed model on the ISIC2018 dataset is 7.6% higher than baseline, 21.5% on the Covid-19 dataset, and 10.4% on the Kvasir dataset. Compared with other advanced models, the HiFuse model performs the best. Our code is open-source and available from https://github.com/huoxiangzuo/HiFuse.
The powerful modeling capabilities of all-attention-based transformer architectures often cause overfitting and - for natural language processing tasks - lead to an implicitly learned internal language model in the autoregressive transformer decoder complicating the integration of external language models. In this paper, we explore relaxed attention, a simple and easy-to-implement smoothing of the attention weights, yielding a two-fold improvement to the general transformer architecture: First, relaxed attention provides regularization when applied to the self-attention layers in the encoder. Second, we show that it naturally supports the integration of an external language model as it suppresses the implicitly learned internal language model by relaxing the cross attention in the decoder. We demonstrate the benefit of relaxed attention across several tasks with clear improvement in combination with recent benchmark approaches. Specifically, we exceed the former state-of-the-art performance of 26.90% word error rate on the largest public lip-reading LRS3 benchmark with a word error rate of 26.31%, as well as we achieve a top-performing BLEU score of 37.67 on the IWSLT14 (DE$\rightarrow$EN) machine translation task without external language models and virtually no additional model parameters. Code and models will be made publicly available.
This paper proposes transferred initialization with modified fully connected layers for COVID-19 diagnosis. Convolutional neural networks (CNN) achieved a remarkable result in image classification. However, training a high-performing model is a very complicated and time-consuming process because of the complexity of image recognition applications. On the other hand, transfer learning is a relatively new learning method that has been employed in many sectors to achieve good performance with fewer computations. In this research, the PyTorch pre-trained models (VGG19\_bn and WideResNet -101) are applied in the MNIST dataset for the first time as initialization and with modified fully connected layers. The employed PyTorch pre-trained models were previously trained in ImageNet. The proposed model is developed and verified in the Kaggle notebook, and it reached the outstanding accuracy of 99.77% without taking a huge computational time during the training process of the network. We also applied the same methodology to the SIIM-FISABIO-RSNA COVID-19 Detection dataset and achieved 80.01% accuracy. In contrast, the previous methods need a huge compactional time during the training process to reach a high-performing model. Codes are available at the following link: github.com/dipuk0506/SpinalNet
Solder joint inspection (SJI) is a critical process in the production of printed circuit boards (PCB). Detection of solder errors during SJI is quite challenging as the solder joints have very small sizes and can take various shapes. In this study, we first show that solders have low feature diversity, and that the SJI can be carried out as a fine-grained image classification task which focuses on hard-to-distinguish object classes. To improve the fine-grained classification accuracy, penalizing confident model predictions by maximizing entropy was found useful in the literature. Inline with this information, we propose using the {\alpha}-skew Jensen-Shannon divergence ({\alpha}-JS) for penalizing the confidence in model predictions. We compare the {\alpha}-JS regularization with both existing entropyregularization based methods and the methods based on attention mechanism, segmentation techniques, transformer models, and specific loss functions for fine-grained image classification tasks. We show that the proposed approach achieves the highest F1-score and competitive accuracy for different models in the finegrained solder joint classification task. Finally, we visualize the activation maps and show that with entropy-regularization, more precise class-discriminative regions are localized, which are also more resilient to noise. Code will be made available here upon acceptance.
Deep convolutional neural networks have shown remarkable performance on various computer vision tasks, and yet, they are susceptible to picking up spurious correlations from the training signal. So called `shortcuts' can occur during learning, for example, when there are specific frequencies present in the image data that correlate with the output predictions. Both high and low frequencies can be characteristic of the underlying noise distribution caused by the image acquisition rather than in relation to the task-relevant information about the image content. Models that learn features related to this characteristic noise will not generalize well to new data. In this work, we propose a simple yet effective training strategy, Frequency Dropout, to prevent convolutional neural networks from learning frequency-specific imaging features. We employ randomized filtering of feature maps during training which acts as a feature-level regularization. In this study, we consider common image processing filters such as Gaussian smoothing, Laplacian of Gaussian, and Gabor filtering. Our training strategy is model-agnostic and can be used for any computer vision task. We demonstrate the effectiveness of Frequency Dropout on a range of popular architectures and multiple tasks including image classification, domain adaptation, and semantic segmentation using both computer vision and medical imaging datasets. Our results suggest that the proposed approach does not only improve predictive accuracy but also improves robustness against domain shift.
Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although convolution neural networks (CNNs) have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph, such as the self-attention operation in Transformers, is beneficial for such modelling, however, its computational overhead is prohibitive. In this paper, we propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. This formulation allows us to design a self-attention module, and more importantly a new Transformer-based backbone network, that we use for both image classification pretraining, and for addressing various downstream tasks (object detection, instance and semantic segmentation). Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on four different tasks. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters. Code and models will be made publicly available at https://github.com/fudan-zvg/DGMN2
We introduce deep tensor networks, which are exponentially wide neural networks based on the tensor network representation of the weight matrices. We evaluate the proposed method on the image classification (MNIST, FashionMNIST) and sequence prediction (cellular automata) tasks. In the image classification case, deep tensor networks improve our matrix product state baselines and achieve 0.49% error rate on MNIST and 8.3% error rate on FashionMNIST. In the sequence prediction case, we demonstrate an exponential improvement in the number of parameters compared to the one-layer tensor network methods. In both cases, we discuss the non-uniform and the uniform tensor network models and show that the latter generalizes well to different input sizes.
Deep neural networks (DNNs) are well known to be vulnerable to adversarial examples (AEs). In addition, AEs have adversarial transferability, which means AEs generated for a source model can fool another black-box model (target model) with a non-trivial probability. In this paper, we investigate the property of adversarial transferability between models including ConvMixer, which is an isotropic network, for the first time. To objectively verify the property of transferability, the robustness of models is evaluated by using a benchmark attack method called AutoAttack. In an image classification experiment, ConvMixer is confirmed to be weak to adversarial transferability.
We propose a new strategy to improve the accuracy and robustness of image classification. First, we train a baseline CNN model. Then, we identify challenging regions in the feature space by identifying all misclassified samples, and correctly classified samples with low confidence values. These samples are then used to train a Variational AutoEncoder (VAE). Next, the VAE is used to generate synthetic images. Finally, the generated synthetic images are used in conjunction with the original labeled images to train a new model in a semi-supervised fashion. Empirical results on benchmark datasets such as STL10 and CIFAR-100 show that the synthetically generated samples can further diversify the training data, leading to improvement in image classification in comparison with the fully supervised baseline approaches using only the available data.
The idea of neural Ordinary Differential Equations (ODE) is to approximate the derivative of a function (data model) instead of the function itself. In residual networks, instead of having a discrete sequence of hidden layers, the derivative of the continuous dynamics of hidden state can be parameterized by an ODE. It has been shown that this type of neural network is able to produce the same results as an equivalent residual network for image classification. In this paper, we design a novel neural ODE for the semantic segmentation task. We start by a baseline network that consists of residual modules, then we use the modules to build our neural ODE network. We show that our neural ODE is able to achieve the state-of-the-art results using 57% less memory for training, 42% less memory for testing, and 68% less number of parameters. We evaluate our model on the Cityscapes, CamVid, LIP, and PASCAL-Context datasets.
Benefited from the rich and detailed spectral information in hyperspectral images (HSI), HSI offers great potential for a wide variety of medical applications such as computational pathology. But, the lack of adequate annotated data and the high spatiospectral dimensions of HSIs usually make classification networks prone to overfit. Thus, learning a general representation which can be transferred to the downstream tasks is imperative. To our knowledge, no appropriate self-supervised pre-training method has been designed for histopathology HSIs. In this paper, we introduce an efficient and effective Self-supervised Spectral Regression (S$^3$R) method, which exploits the low rank characteristic in the spectral domain of HSI. More concretely, we propose to learn a set of linear coefficients that can be used to represent one band by the remaining bands via masking out these bands. Then, the band is restored by using the learned coefficients to reweight the remaining bands. Two pre-text tasks are designed: (1)S$^3$R-CR, which regresses the linear coefficients, so that the pre-trained model understands the inherent structures of HSIs and the pathological characteristics of different morphologies; (2)S$^3$R-BR, which regresses the missing band, making the model to learn the holistic semantics of HSIs. Compared to prior arts i.e., contrastive learning methods, which focuses on natural images, S$^3$R converges at least 3 times faster, and achieves significant improvements up to 14% in accuracy when transferring to HSI classification tasks.
Can we train a hybrid discriminative-generative model within a single network? This question has recently been answered in the affirmative, introducing the field of Joint Energy-based Model (JEM), which achieves high classification accuracy and image generation quality simultaneously. Despite recent advances, there remain two performance gaps: the accuracy gap to the standard softmax classifier, and the generation quality gap to state-of-the-art generative models. In this paper, we introduce a variety of training techniques to bridge the accuracy gap and the generation quality gap of JEM. 1) We incorporate a recently proposed sharpness-aware minimization (SAM) framework to train JEM, which promotes the energy landscape smoothness and the generalizability of JEM. 2) We exclude data augmentation from the maximum likelihood estimate pipeline of JEM, and mitigate the negative impact of data augmentation to image generation quality. Extensive experiments on multiple datasets demonstrate that our SADA-JEM achieves state-of-the-art performances and outperforms JEM in image classification, image generation, calibration, out-of-distribution detection and adversarial robustness by a notable margin. Our code is available at https://github.com/sndnyang/SADAJEM.
The impressive performance of deep learning architectures is associated with a massive increase in model complexity. Millions of parameters need to be tuned, with training and inference time scaling accordingly, together with energy consumption. But is massive fine-tuning always necessary? In this paper, focusing on image classification, we consider a simple transfer learning approach exploiting pre-trained convolutional features as input for a fast-to-train kernel method. We refer to this approach as \textit{top-tuning} since only the kernel classifier is trained on the target dataset. In our study, we perform more than 3000 training processes focusing on 32 small to medium-sized target datasets, a typical situation where transfer learning is necessary. We show that the top-tuning approach provides comparable accuracy with respect to fine-tuning, with a training time between one and two orders of magnitude smaller. These results suggest that top-tuning is an effective alternative to fine-tuning in small/medium datasets, being especially useful when training time efficiency and computational resources saving are crucial.
Adversarial Training (AT), which is commonly accepted as one of the most effective approaches defending against adversarial examples, can largely harm the standard performance, thus has limited usefulness on industrial-scale production and applications. Surprisingly, this phenomenon is totally opposite in Natural Language Processing (NLP) task, where AT can even benefit for generalization. We notice the merit of AT in NLP tasks could derive from the discrete and symbolic input space. For borrowing the advantage from NLP-style AT, we propose Discrete Adversarial Training (DAT). DAT leverages VQGAN to reform the image data to discrete text-like inputs, i.e. visual words. Then it minimizes the maximal risk on such discrete images with symbolic adversarial perturbations. We further give an explanation from the perspective of distribution to demonstrate the effectiveness of DAT. As a plug-and-play technique for enhancing the visual representation, DAT achieves significant improvement on multiple tasks including image classification, object detection and self-supervised learning. Especially, the model pre-trained with Masked Auto-Encoding (MAE) and fine-tuned by our DAT without extra data can get 31.40 mCE on ImageNet-C and 32.77% top-1 accuracy on Stylized-ImageNet, building the new state-of-the-art. The code will be available at https://github.com/alibaba/easyrobust.
Humans learn continually throughout their lifespan by accumulating diverse knowledge and fine-tuning it for future tasks. When presented with a similar goal, neural networks suffer from catastrophic forgetting if data distributions across sequential tasks are not stationary over the course of learning. An effective approach to address such continual learning (CL) problems is to use hypernetworks which generate task dependent weights for a target network. However, the continual learning performance of existing hypernetwork based approaches are affected by the assumption of independence of the weights across the layers in order to maintain parameter efficiency. To address this limitation, we propose a novel approach that uses a dependency preserving hypernetwork to generate weights for the target network while also maintaining the parameter efficiency. We propose to use recurrent neural network (RNN) based hypernetwork that can generate layer weights efficiently while allowing for dependencies across them. In addition, we propose novel regularisation and network growth techniques for the RNN based hypernetwork to further improve the continual learning performance. To demonstrate the effectiveness of the proposed methods, we conducted experiments on several image classification continual learning tasks and settings. We found that the proposed methods based on the RNN hypernetworks outperformed the baselines in all these CL settings and tasks.
The field of continual deep learning is an emerging field and a lot of progress has been made. However, concurrently most of the approaches are only tested on the task of image classification, which is not relevant in the field of intelligent vehicles. Only recently approaches for class-incremental semantic segmentation were proposed. However, all of those approaches are based on some form of knowledge distillation. At the moment there are no investigations on replay-based approaches that are commonly used for object recognition in a continual setting. At the same time while unsupervised domain adaption for semantic segmentation gained a lot of traction, investigations regarding domain-incremental learning in an continual setting is not well-studied. Therefore, the goal of our work is to evaluate and adapt established solutions for continual object recognition to the task of semantic segmentation and to provide baseline methods and evaluation protocols for the task of continual semantic segmentation. We firstly introduce evaluation protocols for the class- and domain-incremental segmentation and analyze selected approaches. We show that the nature of the task of semantic segmentation changes which methods are most effective in mitigating forgetting compared to image classification. Especially, in class-incremental learning knowledge distillation proves to be a vital tool, whereas in domain-incremental learning replay methods are the most effective method.
Mosquito-borne diseases (MBDs), such as dengue virus, chikungunya virus, and West Nile virus, cause over one million deaths globally every year. Because many such diseases are spread by the Aedes and Culex mosquitoes, tracking these larvae becomes critical in mitigating the spread of MBDs. Even as citizen science grows and obtains larger mosquito image datasets, the manual annotation of mosquito images becomes ever more time-consuming and inefficient. Previous research has used computer vision to identify mosquito species, and the Convolutional Neural Network (CNN) has become the de-facto for image classification. However, these models typically require substantial computational resources. This research introduces the application of the Vision Transformer (ViT) in a comparative study to improve image classification on Aedes and Culex larvae. Two ViT models, ViT-Base and CvT-13, and two CNN models, ResNet-18 and ConvNeXT, were trained on mosquito larvae image data and compared to determine the most effective model to distinguish mosquito larvae as Aedes or Culex. Testing revealed that ConvNeXT obtained the greatest values across all classification metrics, demonstrating its viability for mosquito larvae classification. Based on these results, future research includes creating a model specifically designed for mosquito larvae classification by combining elements of CNN and transformer architecture.
Gastro-Intestinal Tract cancer is considered a fatal malignant condition of the organs in the GI tract. Due to its fatality, there is an urgent need for medical image segmentation techniques to segment organs to reduce the treatment time and enhance the treatment. Traditional segmentation techniques rely upon handcrafted features and are computationally expensive and inefficient. Vision Transformers have gained immense popularity in many image classification and segmentation tasks. To address this problem from a transformers' perspective, we introduced a hybrid CNN-transformer architecture to segment the different organs from an image. The proposed solution is robust, scalable, and computationally efficient, with a Dice and Jaccard coefficient of 0.79 and 0.72, respectively. The proposed solution also depicts the essence of deep learning-based automation to improve the effectiveness of the treatment
Recent advances in artificial intelligence promote a wide range of computer vision applications in many different domains. Digital cameras, acting as human eyes, can perceive fundamental object properties, such as shapes and colors, and can be further used for conducting high-level tasks, such as image classification, and object detections. Human perceptions have been widely recognized as the ground truth for training and evaluating computer vision models. However, in some cases, humans can be deceived by what they have seen. Well-functioned human vision relies on stable external lighting while unnatural illumination would influence human perception of essential characteristics of goods. To evaluate the illumination effects on human and computer perceptions, the group presents a novel dataset, the Food Vision Dataset (FVD), to create an evaluation benchmark to quantify illumination effects, and to push forward developments of illumination estimation methods for fair and reliable consumer acceptability prediction from food appearances. FVD consists of 675 images captured under 3 different power and 5 different temperature settings every alternate day for five such days.
In recent years novel architecture components for image classification have been developed, starting with attention and patches used in transformers. While prior works have analyzed the influence of some aspects of architecture components on the robustness to adversarial attacks, in particular for vision transformers, the understanding of the main factors is still limited. We compare several (non)-robust classifiers with different architectures and study their properties, including the effect of adversarial training on the interpretability of the learnt features and robustness to unseen threat models. An ablation from ResNet to ConvNeXt reveals key architectural changes leading to almost $10\%$ higher $\ell_\infty$-robustness.