331 research outputs found
Managing diversity, innovation, and infrastructure in digital business/ Nilanjan Ray, editor.
No full text"This book presents a variety of practical applications, tools, skills, practices, models, approaches, and strategies that are proving themselves in practice-demonstrating effectiveness with managing diversity and innovation. It also shares a different pattern of research work that will provide a platform for new avenues in overall infrastructure, socioeconomic conditions, and modern tourism business infrastructure"--1 online resource
Unsupervised domain adaptation for object detection and whole slide image classification
No full textDeep neural network (DNN) has been developed rapidly in years. While it shows promising results in various tasks of computer vision, DNN typically suffers from accuracy loss due to the domain shift from a source domain to a target domain. To mitigate the accuracy loss without the label from target domain, unsupervised domain adaptation (UDA) approaches are proposed.
Compare to most UDA studies that target image classification and pixel-level classification (image segmentation), UDA for object detection is a relatively new area. A popular processing pipeline is to apply adversarial training with domain discriminator. The domain discriminator aligns the feature distributions of the source and target domain.
Existing methods in UDA object detection extract features from image level and directly adapt the full features as in UDA for classification tasks. However, alignment on full image level features as a whole is not ideal for object detection task. The presence of varied backgrounds could interfere with the result of adaptation. To avoid alignment on a full feature, this thesis proposes a novel foreground-focused domain adaptation (FFDA) framework. This FFDA framework mines the loss of the domain discriminators so that the alignment could concentrate on the foreground during backpropagation.
FFDA collects target predictions and source image labels and uses them to generate mining masks that outline foreground regions. And then it applies the masks to image and instance level domain discriminators to allow backpropagation only on mined regions. In addition, by reinforcing this foreground-focused adaptation throughout multiple layers in the detector model, FFDA pushes the detector to gain a significant accuracy boost on target domain prediction. Compared with previous methods, FFDA method reaches the new state-of-the-art accuracy on adaptation from Cityscape to Foggy Cityscape dataset. The FFDA also demonstrates competitive results on other datasets that include various scenarios for autonomous driving applications.
In addition to object detection problem, this thesis also discusses the application of UDA for whole slide image (WSI) classification. Image classification for WSI is a challenging task compared to general image classification because of its high resolution and scattered key information. Previous work provided a novel deep Fisher vector coding pipeline for WSI classification. However, this pipeline suffers from the same accuracy drop phenomenon when deployed to another set of WSI from a different institution to perform the same task. This poses a limitation of the practical usage of the pipeline especially when the diagnoses of WSIs are hard to obtain.
On the other hand, previous works that apply UDA to medical imaging typically focused on adapting on small microscopy image samples or image patches extracted from WSI. UDA for the application of classifying the entire WSI has not yet been discussed due to the limited number of pipelines and datasets that support WSI classification.
This thesis aims at providing a UDA solution to enhance the robustness of the previous pipeline by mitigating the accuracy drop caused by different WSI datasets. This solution inserts the domain classifiers into the previous pipeline in different stages to align the features during training. The solution is evaluated by calculating confusion matrices before and after the adaptation. The results demonstrate that by placing domain classifiers in different stages the pipeline shows an accuracy boost on target WSI data
Mixed Low-bit Quantization for Model Compression with Layer Importance and Gradient Estimations
No full textDeep neural networks (DNNs) have been widely used in the modern world in recent years. However, due to the substantial memory consumption and high computational power use of DNNs, deploying them on devices with limited resources is challenging. Model compression methods can provide us with a remedy here. Among those techniques, neural network quantization has achieved a high compression rate using a low bitwidth representation of weights and activations while maintaining the accuracy of the high-precision original network. However, mixed precision (per-layer bit-width precision) quantization requires careful tuning to maintain accuracy while achieving further compression and higher granularity than fixed precision quantization. In this thesis, We propose an accuracy-aware criterion to quantify the layer’s importance rank. Our method applies imprinting per layer, which acts as a proxy module for accuracy estimation in an efficient way. We rank the layers based on the accuracy gain from previous modules and iteratively quantize those with less accuracy. Previous mixed-precision methods either rely on expensive search techniques such as reinforcement learning (RL) or end-to-end optimization with a lack of interpretation
to the quantization configuration scheme. Our method is a one-shot, efficient, accuracy-aware information estimation and thus draws better interpretability to the selected bit-width configuration. We have also pointed out the problem of
the Straight-Through Estimator (STE), which is commonly used for gradients estimation in the quantization field. We’ve discussed some ways to address the problem of using STE
A Closer Look at Weak Supervision’s Limitations in WSI Recurrence Score Prediction
No full textHistological examination and derived ancillary testing remain the gold standard for breast cancer diagnosis, prognosis assessment and treatment guidance. Currently, a commercial molecular signature test OncotypeDX®, based on RNA quantitation and providing a recurrence score (RS) ranging from 0 to 100, is routinely utilized for luminal breast cancers (the largest sub-type group of breast cancers) to predict the probabilities of response to chemotherapy and disease recurrence. We attempt to predict RS using digital pathology and Weakly Supervised (WS) attention-based models. In tissue samples, the malignant component is haphazardly admixed with the non-malignant component in variable proportions. This represents a challenge for WS attention-based models to identify high-valued diagnostic/prognostic areas within whole slide images (WSIs). To address this, we propose an interactive, supervised approach with a human in the middle by creating a user-friendly Graphical User Interface (GUI) that allows an expert pathologist to annotate heatmaps generated by any WS attention-based model. We aim to enhance the model’s learning capabilities and performance by incorporating the feedback from the GUI as expected scores in the successive training process. We train WS attention-based models like CLAM (Clustering-constrained Attention Multiple Instance Learning) and TransMIL (Transformer based Correlated Multiple Instance Learning) on our in-house dataset before and after the expert feedback. We observe an improvement in RS prediction after retraining both models with the pathologist’s annotation- a 5% rise in validation-test AUC and 4% in validation-test accuracy for CLAM and a 4.5% increase in validation-test AUC and 3% in validation-test accuracy for TransMIL. We analyze the generated heatmaps and observe how additional supervision from a domain expert enhances the learning capacity of the models. We notice an improvement in cosine similarity between the pathologist’s GUI-based attention scores and trained models’ attention maps after feedback - 5% and 10% increase for CLAM and TransMIL, respectively. Our adaptive, interactive system harmonizes attention scores with expert intuition and instills higher confidence in the system’s predictions. This study establishes a potent synergy between AI and expert collaboration, addressing the constraints of WS by enhancing the discrimination of diagnostic features and making an effort to generate predictions according to clinical diagnostic norms
End-to-End Learning of Dynamic Programming and Convolutional Neural Networks using Differentiable Bypass
No full textDifferentiable Programming is the paradigm where different functions or modules are combined into a unified pipeline with the purpose of applying end-to-end learning or optimization. A natural impediment is the non-differentiability characteristic of many modules. This thesis proposes a new way to overcome this obstacle by using a concept called Differentiable Bypass (DiffBypass). DiffBypass exploits the Universal Function Approximation property of neural networks to mimic the output of non-differentiable functions or modules in the pipeline, rerouting the gradient path to bypass these components entirely.Further, as a significant application, we demonstrate the use of DiffBypass to combine Convolutional Neural Networks (CNN) and Dynamic Programming (DP) in end-to-end learning for segmenting left ventricle from short axis view of heart MRI. Our experiments show that end-to-end combination of CNN and DP requires fewer labeled images to achieve a significantly better segmentation accuracy than using only CNN by allowing the incorporation of strong prior knowledge into the pipeline to cope with lack of training data. Comparison between DiffBypass and Evolution Strategy (ES), another method that can be used to train non-differentiable modules, shows that DiffBypass is more robust and has better performance for high-dimension problems.Finally, as a technical contribution, we provide a set of recommendations for training non-differentiable modules using DiffBypass. Furthermore, we also provide a code base for reproducibility. We think DiffBypass has the potential to become a blueprint to expand differentiable programming to include non-differentiable modules
Moving Object Detection Using Unsupervised and Weakly Supervised Neural Networks in Videos with Illumination Changes and Dynamic Background
No full textBackground subtraction is a crucial task in computer vision applications, such as video surveillance, traffic monitoring, autonomous navigation, and human-computer interaction. This approach involves acquiring a background model to separate moving objects and the background from an input image. However, challenges such as sudden and gradual illumination changes and dynamic backgrounds can make this task difficult. Among various methods proposed for background subtraction, supervised deep learning-based techniques are currently considered state-of-the-art. However, these methods require pixel-wise ground-truth labeling, which can be time-consuming and expensive. The aim of this thesis is to develop unsupervised and weakly supervised background subtraction methods that can handle illumination changes or dynamic backgrounds.
Most methods handle illumination changes and shadows in batch mode and are unsuitable for long video sequences or real-time applications. To address this, we propose an extension of a state-of-the-art batch Moving Object Detection (MOD) method, ILISD, to an online/incremental MOD method using unsupervised and generative neural networks. Our method uses illumination invariant image representations and obtains a low-dimensional representation of the background image using a neural network. It then decomposes the foreground image into illumination changes and moving objects.
Yet another challenge is dynamic background, where a background pixel can have different values due to periodical or irregular movements, negatively affecting a method's performance. For example, surging of water, water fountains and waving trees cause dynamic variations in the background. To address this, we propose a new unsupervised method, called DBSGen, which estimates a dense dynamic motion map using a generative multi-resolution convolutional network and warps the input images by the obtained motion map. Then, a generative fully connected network generates background images using the warped input images in its reconstruction loss term. Finally, a pixel-wise distance threshold that utilizes a dynamic entropy map obtains the binary segmented results.
Finally, we propose a weakly supervised background subtraction method, where the training set consists of a moving object-free sequence of images, without requiring per-pixel ground-truth annotations. Our method consists of two neural networks. The first network, an autoencoder, generates dynamic background images for training the second network. Dynamic background images are obtained by applying a threshold to background-subtracted images. The second network is a U-Net that uses as input the same moving object-free video and is trained by using the dynamic background images produced by the autoencoder for its target output. During the testing phase, the autoencoder and U-Net process input images to generate background and dynamic background images, respectively. The dynamic background image helps remove dynamic motion from the background-subtracted image, resulting in a foreground image that is free of dynamic artifacts
Geographical Patterns in the Tourist City:GIS for Spatiotemporal Analysis
No full textUnderstanding how tourists use space in urban environments has practical applications in destination management and planning. It reveals key issues for organizing facilities and essential services, creating new products, impact management, and other purposes related to urban tourism. The geotagged photos shared by Panoramio users, during their visit to the city of Lisbon between 2007 and 2014, allow the authors to incorporate into this chapter a quantitative and geographic reading on the tourist consumption of spaces within the city. In this chapter, they analyze the spatial distribution of tourists and its changes across time considering a period of 8 years. Furthermore, a regression analysis method was carried out in order to find the spatial relations between the observed pattern (geographical agglomeration of tourist ́s photos) and a set of 24 selected independent variables. </jats:p
Learning Instance-Specific Parameters of Black-Box Models Using Differentiable Surrogates
No full textTuning parameters of a non-differentiable or black-box compute is challenging. Existing methods rely mostly on random sampling or grid sampling from the parameter space. Further, with all the current methods, it is not possible to supply any input specific parameters to the black-box. To the best of our knowledge, for the first time, we are able to learn input-specific parameters for a black box in this work. As a test application, we choose a popular image denoising method BM3D as our black-box compute. Then, we use a differentiable surrogate model (a neural network) to approximate the black-box behaviour. Next, another neural network is used in an end-to-end fashion to learn input instance-specific parameters for the black-box. Motivated by prior advances in surrogate-based optimization, we applied our method to the Smartphone Image Denoising Dataset (SIDD) and the Color Berkeley Segmentation Dataset (CBSD68) for image denoising. The results are compelling, demonstrating a significant increase in PSNR and a notable improvement in SSIM nearing 0.93. Experimental results underscore the effectiveness of our approach in achieving substantial improvements in both model performance and optimization efficiency. For code and implementation details, please refer to our GitHub repository:
https://github.com/arnisha-k/instance-specific-param
Unknown-box Approximation to Improve Optical Character Recognition Performance
No full textOptical character recognition (OCR) is a widely used pattern recognition application in numerous domains. Several feature-rich commercial OCR solutions and opensource OCR solutions are available for consumers, which can provide moderate to excellent accuracy levels. These solutions are general-purpose by design to serve a wider community. However, accuracy can diminish with difficult and uncommon document domains. Preprocessing of document images can be used to minimize the effect of domain shift. In this thesis, we investigate the possibility and the effect of using OCR engine feedback to train a preprocessor. The main obstacle in this approach is propagating the error signal through an opaque OCR engine. Circumventing this obstacle, we propose a novel preprocessor trained using gradient approximation. Unlike the previous OCR agnostic preprocessing techniques, the proposed training approach approximates a particular OCR engine's gradient and trains the preprocessor module eliminating the need for intermediate labels. We compare two different methods to our proposed approach to establish a better training pipeline. Experiments with two different datasets and two OCR engines show that the presented preprocessor is able to improve the accuracy of the OCR engine from the baseline accuracy by applying pixel-level manipulations to the document image
Unpaired Document Image Denoising for OCR using BiLSTM enhanced CycleGAN
No full textThe recognition performance of Optical Character Recognition (OCR) models can be sub-optimal when document images suffer from various degradations. Supervised learning-based methods for image enhancement can generate high-quality enhanced images. However, these methods require the availability of corresponding clean images or ground truth text for training. Moreover, the paired training data used for training these models is usually generated by adding different types of synthetic noise to clean images. Real-world noise is more challenging and complex in nature compared to synthetic noise. To effectively enhance real-world noisy images, the models must be trained using real noisy images. However, it is infeasible to have corresponding clean images for real-world noisy images, and creating ground truth text requires manual effort. Unsupervised methods have been explored in recent years, focusing on enhancing natural scene images. In the case of document images, preserving the readability of text in the enhanced images is of utmost importance for improved OCR performance. In this thesis, we explore the possibility of enhancing documents in an unsupervised setting using unpaired training samples. To this end, we propose a modified architecture for the standard CycleGAN model to improve its performance in enhancing document images with better text preservation. The results indicate that the proposed model leads to better preservation of text and improved OCR performance compared to the CycleGAN model and classical unsupervised image preprocessing techniques like Sauvola and Otsu
- …
