Deep neural networks (DNNs) achieve excellent performance on standard classification tasks. However, under image quality distortions such as blur and noise, classification accuracy becomes poor. In this work, we compare the performance of DNNs with human subjects on distorted images. We show that, although DNNs perform better than or on par with humans on good quality images, DNN performance is still much lower than human performance on distorted images. We additionally find that there is little correlation in errors between DNNs and human subjects. This could be an indication that the internal representation of images are different between DNNs and the human visual system. These comparisons with human performance could be used to guide future development of more robust DNNs.
Deep learning architectures are showing great promise in various computer vision domains including image classification, object detection, event detection and action recognition. In this study, we investigate various aspects of convolutional neural networks (CNNs) from the big data perspective. We analyze recent studies and different network architectures both in terms of running time and accuracy. We present extensive empirical information along with best practices for big data practitioners. Using these best practices we propose efficient fusion mechanisms both for single and multiple network models. We present state-of-the art results on benchmark datasets while keeping computational costs at a lower level. Another contribution of our paper is that these state-of-the-art results can be reached without using extensive data augmentation techniques.
Rotation moment invariants have been of great interest in image processing and pattern recognition. This paper presents a novel kind of rotation moment invariants based on the Slepian functions, which were originally introduced in the method of separation of variables for Helmholtz equations. They were first proposed for time series by Slepian and his coworkers in the 1960s. Recent studies have shown that these functions have an good performance in local approximation compared to other approximation basis. Motivated by the good approximation performance, we construct the Slepian-based moments and derive the rotation invariant. We not only theoretically prove the invariance, but also discuss the experiments on real data. The proposed rotation invariants are robust to noise and yield decent performance in facial expression classification.
We demonstrate in this paper that a generative model can be designed to perform classification tasks under challenging settings, including adversarial attacks and input distribution shifts. Specifically, we propose a conditional variational autoencoder that learns both the decomposition of inputs and the distributions of the resulting components. During test, we jointly optimize the latent variables of the generator and the relaxed component labels to find the best match between the given input and the output of the generator. The model demonstrates promising performance at recognizing overlapping components from the multiMNIST dataset, and novel component combinations from a traffic sign dataset. Experiments also show that the proposed model achieves high robustness on MNIST and NORB datasets, in particular for high-strength gradient attacks and non-gradient attacks.
Identifying the origin of a sample image in biometric systems can be beneficial for data authentication in case of attacks against the system and for initiating sensor-specific processing pipelines in sensor-heterogeneous environments. Motivated by shortcomings of the photo response non-uniformity (PRNU) based method in the biometric context, we use a texture classification approach to detect the origin of finger vein sample images. Based on eight publicly available finger vein datasets and applying eight classical yet simple texture descriptors and SVM classification, we demonstrate excellent sensor model identification results for raw finger vein samples as well as for the more challenging region of interest data. The observed results establish texture descriptors as effective competitors to PRNU in finger vein sensor model identification.
Analysis of cardiac ultrasound images is commonly performed in routine clinical practice for quantification of cardiac function. Its increasing automation frequently employs deep learning networks that are trained to predict disease or detect image features. However, such models are extremely data-hungry and training requires labelling of many thousands of images by experienced clinicians. Here we propose the use of contrastive learning to mitigate the labelling bottleneck. We train view classification models for imbalanced cardiac ultrasound datasets and show improved performance for views/classes for which minimal labelled data is available. Compared to a naive baseline model, we achieve an improvement in F1 score of up to 26% in those views while maintaining state-of-the-art performance for the views with sufficiently many labelled training observations.
Disease diagnosis from medical images via supervised learning is usually dependent on tedious, error-prone, and costly image labeling by medical experts. Alternatively, semi-supervised learning and self-supervised learning offer effectiveness through the acquisition of valuable insights from readily available unlabeled images. We present Semi-Supervised Relational Contrastive Learning (SRCL), a novel semi-supervised learning model that leverages self-supervised contrastive loss and sample relation consistency for the more meaningful and effective exploitation of unlabeled data. Our experimentation with the SRCL model explores both pre-train/fine-tune and joint learning of the pretext (contrastive learning) and downstream (diagnostic classification) tasks. We validate against the ISIC 2018 Challenge benchmark skin lesion classification dataset and demonstrate the effectiveness of our semi-supervised method on varying amounts of labeled data.
We revisit the "dataset classification" experiment suggested by Torralba and Efros a decade ago, in the new era with large-scale, diverse, and hopefully less biased datasets as well as more capable neural network architectures. Surprisingly, we observe that modern neural networks can achieve excellent accuracy in classifying which dataset an image is from: e.g., we report 84.7% accuracy on held-out validation data for the three-way classification problem consisting of the YFCC, CC, and DataComp datasets. Our further experiments show that such a dataset classifier could learn semantic features that are generalizable and transferable, which cannot be simply explained by memorization. We hope our discovery will inspire the community to rethink the issue involving dataset bias and model capabilities.
Deep learning for object classification relies heavily on convolutional models. While effective, CNNs are rarely interpretable after the fact. An attention mechanism can be used to highlight the area of the image that the model focuses on thus offering a narrow view into the mechanism of classification. We expand on this idea by forcing the method to explicitly align images to be classified to reference images representing the classes. The mechanism of alignment is learned and therefore does not require that the reference objects are anything like those being classified. Beyond explanation, our exemplar based cross-alignment method enables classification with only a single example per category (one-shot). Our model cuts the 5-way, 1-shot error rate in Omniglot from 2.1% to 1.4% and in MiniImageNet from 53.5% to 46.5% while simultaneously providing point-wise alignment information providing some understanding on what the network is capturing. This method of alignment also enables the recognition of an unsupported class (open-set) in the one-shot setting while maintaining an F1-score of above 0.5 for Omniglot even with 19 other distracting classes while baselines completely fail to separate the open-set class in the one-shot setting.
Recent few-shot learning works focus on training a model with prior meta-knowledge to fast adapt to new tasks with unseen classes and samples. However, conventional time-series classification algorithms fail to tackle the few-shot scenario. Existing few-shot learning methods are proposed to tackle image or text data, and most of them are neural-based models that lack interpretability. This paper proposes an interpretable neural-based framework, namely \textit{Dual Prototypical Shapelet Networks (DPSN)} for few-shot time-series classification, which not only trains a neural network-based model but also interprets the model from dual granularity: 1) global overview using representative time series samples, and 2) local highlights using discriminative shapelets. In particular, the generated dual prototypical shapelets consist of representative samples that can mostly demonstrate the overall shapes of all samples in the class and discriminative partial-length shapelets that can be used to distinguish different classes. We have derived 18 few-shot TSC datasets from public benchmark datasets and evaluated the proposed method by comparing with baselines. The DPSN framework outperforms state-of-the-art time-series classification methods, especially when training with limited amounts of data. Several case studies have been given to demonstrate the interpret ability of our model.
In this study we investigate the potential of parametric images formed from ultrasound B-mode scans using the Nakagami distribution for non-invasive classification of breast lesions. Through a sliding window technique, we generated seven types of parametric images from each patient scan in our dataset using basic and as well as derived parameters of the Nakagami distribution. To determine the most suitable window size for image generation, we conducted an empirical analysis using three windows, and selected the best one for our study. From the parametric images formed for each patient, we extracted a total of 72 features. Feature selection was performed to find the optimum subset of features for the best classification performance. Incorporating the selected subset of features with the Support Vector Machine (SVM) classifier, and by tuning the decision threshold, we obtained a maximum classification accuracy of 93.08%, an Area under the ROC Curve (AUC) of 0.9712, a False Negative Rate of 0%, and a very low False Positive Rate of 8.65%. Our results indicate that the high accuracy of such a procedure may assist in the diagnostic process associated with detection of breast cancer, as well as help to reduce false positive diagnosis.
Published research highlights the presence of demographic bias in automated facial attribute classification. The proposed bias mitigation techniques are mostly based on supervised learning, which requires a large amount of labeled training data for generalizability and scalability. However, labeled data is limited, requires laborious annotation, poses privacy risks, and can perpetuate human bias. In contrast, self-supervised learning (SSL) capitalizes on freely available unlabeled data, rendering trained models more scalable and generalizable. However, these label-free SSL models may also introduce biases by sampling false negative pairs, especially at low-data regimes 200K images) under low compute settings. Further, SSL-based models may suffer from performance degradation due to a lack of quality assurance of the unlabeled data sourced from the web. This paper proposes a fully self-supervised pipeline for demographically fair facial attribute classifiers. Leveraging completely unlabeled data pseudolabeled via pre-trained encoders, diverse data curation techniques, and meta-learning-based weighted contrastive learning, our method significantly outperforms existing SSL approaches proposed for downstream image classification tasks. Extensive evaluations on the FairFace and CelebA datasets demonstrate the efficacy of our pipeline in obtaining fair performance over existing baselines. Thus, setting a new benchmark for SSL in the fairness of facial attribute classification.
Many of our core assumptions about how neural networks operate remain empirically untested. One common assumption is that convolutional neural networks need to be stable to small translations and deformations to solve image recognition tasks. For many years, this stability was baked into CNN architectures by incorporating interleaved pooling layers. Recently, however, interleaved pooling has largely been abandoned. This raises a number of questions: Are our intuitions about deformation stability right at all? Is it important? Is pooling necessary for deformation invariance? If not, how is deformation invariance achieved in its absence? In this work, we rigorously test these questions, and find that deformation stability in convolutional networks is more nuanced than it first appears: (1) Deformation invariance is not a binary property, but rather that different tasks require different degrees of deformation stability at different layers. (2) Deformation stability is not a fixed property of a network and is heavily adjusted over the course of training, largely through the smoothness of the convolutional filters. (3) Interleaved pooling layers are neither necessary nor sufficient for achieving the optimal form of deformation stability for natural image classification. (4) Pooling confers too much deformation stability for image classification at initialization, and during training, networks have to learn to counteract this inductive bias. Together, these findings provide new insights into the role of interleaved pooling and deformation invariance in CNNs, and demonstrate the importance of rigorous empirical testing of even our most basic assumptions about the working of neural networks.
Recently, recognition of gender from facial images has gained a lot of importance. There exist a handful of research work that focus on feature extraction to obtain gender specific information from facial images. However, analyzing different facial regions and their fusion help in deciding the gender of a person from facial images. In this paper, we propose a new approach to identify gender from frontal facial images that is robust to background, illumination, intensity, and facial expression. In our framework, first the frontal face image is divided into a number of distinct regions based on facial landmark points that are obtained by the Chehra model proposed by Asthana et al. The model provides 49 facial landmark points covering different regions of the face, e.g. forehead, left eye, right eye, lips. Next, a face image is segmented into facial regions using landmark points and features are extracted from each region. The Compass LBP feature, a variant of LBP feature, has been used in our framework to obtain discriminative gender-specific information. Following this, a Support Vector Machine based classifier has been used to compute the probability scores from each facial region. Finally, the classification scores obtained from individual regions are combined with a genetic algorithm based learning to improve the overall classification accuracy. The experiments have been performed on popular face image datasets such as Adience, cFERET (color FERET), LFW and two sketch datasets, namely CUFS and CUFSF. Through experiments, we have observed that, the proposed method outperforms existing approaches.
Despite remarkable advances in automated visual recognition by machines, some visual tasks remain challenging for machines. Fleuret et al. (2011) introduced the Synthetic Visual Reasoning Test (SVRT) to highlight this point, which required classification of images consisting of randomly generated shapes based on hidden abstract rules using only a few examples. Ellis et al. (2015) demonstrated that a program synthesis approach could solve some of the SVRT problems with unsupervised, few-shot learning, whereas they remained challenging for several convolutional neural networks trained with thousands of examples. Here we re-considered the human and machine experiments, because they followed different protocols and yielded different statistics. We thus proposed a quantitative reintepretation of the data between the protocols, so that we could make fair comparison between human and machine performance. We improved the program synthesis classifier by correcting the image parsings, and compared the results to the performance of other machine agents and human subjects. We grouped the SVRT problems into different types by the two aspects of the core characteristics for classification: shape specification and location relation. We found that the program synthesis classifier could not solve problems involving shape distances, because it relied on symbolic computation which scales poorly with input dimension and adding distances into such computation would increase the dimension combinatorially with the number of shapes in an image. Therefore, although the program synthesis classifier is capable of abstract reasoning, its performance is highly constrained by the accessible information in image parsings.
Distance metric learning (DML) is a critical factor for image analysis and pattern recognition. To learn a robust distance metric for a target task, we need abundant side information (i.e., the similarity/dissimilarity pairwise constraints over the labeled data), which is usually unavailable in practice due to the high labeling cost. This paper considers the transfer learning setting by exploiting the large quantity of side information from certain related, but different source tasks to help with target metric learning (with only a little side information). The state-of-the-art metric learning algorithms usually fail in this setting because the data distributions of the source task and target task are often quite different. We address this problem by assuming that the target distance metric lies in the space spanned by the eigenvectors of the source metrics (or other randomly generated bases). The target metric is represented as a combination of the base metrics, which are computed using the decomposed components of the source metrics (or simply a set of random bases); we call the proposed method, decomposition-based transfer DML (DTDML). In particular, DTDML learns a sparse combination of the base metrics to construct the target metric by forcing the target metric to be close to an integration of the source metrics. The main advantage of the proposed method compared with existing transfer metric learning approaches is that we directly learn the base metric coefficients instead of the target metric. To this end, far fewer variables need to be learned. We therefore obtain more reliable solutions given the limited side information and the optimization tends to be faster. Experiments on the popular handwritten image (digit, letter) classification and challenge natural image annotation tasks demonstrate the effectiveness of the proposed method.
ImageNet has been arguably the most popular image classification benchmark, but it is also the one with a significant level of label noise. Recent studies have shown that many samples contain multiple classes, despite being assumed to be a single-label benchmark. They have thus proposed to turn ImageNet evaluation into a multi-label task, with exhaustive multi-label annotations per image. However, they have not fixed the training set, presumably because of a formidable annotation cost. We argue that the mismatch between single-label annotations and effectively multi-label images is equally, if not more, problematic in the training setup, where random crops are applied. With the single-label annotations, a random crop of an image may contain an entirely different object from the ground truth, introducing noisy or even incorrect supervision during training. We thus re-label the ImageNet training set with multi-labels. We address the annotation cost barrier by letting a strong image classifier, trained on an extra source of data, generate the multi-labels. We utilize the pixel-wise multi-label predictions before the final pooling layer, in order to exploit the additional location-specific supervision signals. Training on the re-labeled samples results in improved model performances across the board. ResNet-50 attains the top-1 classification accuracy of 78.9% on ImageNet with our localized multi-labels, which can be further boosted to 80.2% with the CutMix regularization. We show that the models trained with localized multi-labels also outperforms the baselines on transfer learning to object detection and instance segmentation tasks, and various robustness benchmarks. The re-labeled ImageNet training set, pre-trained weights, and the source code are available at {https://github.com/naver-ai/relabel_imagenet}.
Scene labeling task is to segment the image into meaningful regions and categorize them into classes of objects which comprised the image. Commonly used methods typically find the local features for each segment and label them using classifiers. Afterward, labeling is smoothed in order to make sure that neighboring regions receive similar labels. However, they ignore expressive and non-local dependencies among regions due to expensive training and inference. In this paper, we propose to use high-level knowledge regarding rules in the inference to incorporate dependencies among regions in the image to improve scores of classification. Towards this aim, we extract these rules from data and transform them into constraints for Integer Programming to optimize the structured problem of assigning labels to super-pixels (consequently pixels) of an image. In addition, we propose to use soft-constraints in some scenarios, allowing violating the constraint by imposing a penalty, to make the model more flexible. We assessed our approach on three datasets and obtained promising results.
We present an image translation approach to generate augmented data for mitigating data imbalances in a dataset of histopathology images of colorectal polyps, adenomatous tumors that can lead to colorectal cancer if left untreated. By applying cycle-consistent generative adversarial networks (CycleGANs) to a source domain of normal colonic mucosa images, we generate synthetic colorectal polyp images that belong to diagnostically less common polyp classes. Generated images maintain the general structure of their source image but exhibit adenomatous features that can be enhanced with our proposed filtration module, called Path-Rank-Filter. We evaluate the quality of generated images through Turing tests with four gastrointestinal pathologists, finding that at least two of the four pathologists could not identify generated images at a statistically significant level. Finally, we demonstrate that using CycleGAN-generated images to augment training data improves the AUC of a convolutional neural network for detecting sessile serrated adenomas by over 10%, suggesting that our approach might warrant further research for other histopathology image classification tasks.
We propose a novel non-parametric method for cross-modal recipe retrieval which is applied on top of precomputed image and text embeddings. By combining our method with standard approaches for building image and text encoders, trained independently with a self-supervised classification objective, we create a baseline model which outperforms most existing methods on a challenging image-to-recipe task. We also use our method for comparing image and text encoders trained using different modern approaches, thus addressing the issues hindering the development of novel methods for cross-modal recipe retrieval. We demonstrate how to use the insights from model comparison and extend our baseline model with standard triplet loss that improves state-of-the-art on the Recipe1M dataset by a large margin, while using only precomputed features and with much less complexity than existing methods. Further, our approach readily generalizes beyond recipe retrieval to other challenging domains, achieving state-of-the-art performance on Politics and GoodNews cross-modal retrieval tasks.