Learning identity-Aware, domain-invariant representations is crucial in solving domain generalizable person ReID (DG-ReID). Existing methods commonly use augmentation techniques either in feature space by mixing insta...
详细信息
Implicit neural representation (INR) has emanated as a powerful paradigm for 2D image representation. Recent works like INR-GAN have successfully adopted INR for 2D image synthesis. However, these lack explicit contro...
详细信息
ISBN:
(纸本)9781450398220
Implicit neural representation (INR) has emanated as a powerful paradigm for 2D image representation. Recent works like INR-GAN have successfully adopted INR for 2D image synthesis. However, these lack explicit control on the generated images as achieved by their 3D-aware image synthesis counterparts like GIRAFFE. Our work investigates INRs for the task of controllable image synthesis. We propose a novel framework that allows for manipulation of foreground, background and their shape and appearance in the latent space. To achieve effective control over these attributes, we introduce a novel feature mask coupling technique that leverages the foreground and background masks for mutual learning. Extensive quantitative and qualitative analysis shows that our model can disentangle the latent space successfully and allows to change the foreground and/or background’s shape and appearance. We further demonstrate that our network takes lesser training time than other INR-based image synthesis methods.
Usually, image binarization plays a crucial role in automatic analysis of degraded documents from their captured images. However, this binarization task is often difficult due to a number of reasons including the high...
详细信息
ISBN:
(纸本)9781450398220
Usually, image binarization plays a crucial role in automatic analysis of degraded documents from their captured images. However, this binarization task is often difficult due to a number of reasons including the high similarity between noisy background and faded foreground pixels. The study presented here is particularly focused on binarization of images of low-resource degraded quality documents based on a set of recently collected image samples of several rare, ancient and severely degraded quality printed documents of Bangla, the 2nd and 5th most popular script of India and the world respectively. This new collection of degraded document image samples will henceforth be referred as ’ISIDDI2’ and it consists of 139 images of Bangla old document pages. Samples of ’ISIDDI’, another existing database of degraded Bangla document image samples, have also been used in the present study. A novel deep architecture based on attention UNET++ with dilated convolution operation is proposed for this binarization task. The model is optimized using human vision perceptible distance reciprocal distortion (DRD) loss. Since the binarization ground truth of samples of both ’ISIDDI2’ and ’ISIDDI’ are not available, the proposed network has been trained using samples of DIBCO and H-DIBCO datasets and an unsupervised domain adaptation (DA) module is employed for adaptation of the proposed architecture to the degradation patterns of ’ISIDDI2’ or ’ISIDDI’ samples. The proposed binarization strategy includes certain post-processing operation based on a modified k-neighbourhood based approach for recovery of broken characters. Results of our extensive experimentation show that the proposed binarization strategy has improved the binarization output of state-of-the-art methods on both ISIDDI2 and ISIDDI datasets. Also, its performance on well-known DIBCO samples is satisfactory.
In excision biopsy, a tumor mass is surgically removed from the body. Subsequently, it is sliced at an appropriate location and investigated microscopically through a process called histopathology. Any bias in tumor s...
详细信息
Dehazing is a difficult process because of the damage caused by the non-uniform fog and haze distribution in images. To address these issues, a Multi-Scale Residual dense Dehazing Network (MSRDNet) is proposed in this...
详细信息
ISBN:
(纸本)9781450398220
Dehazing is a difficult process because of the damage caused by the non-uniform fog and haze distribution in images. To address these issues, a Multi-Scale Residual dense Dehazing Network (MSRDNet) is proposed in this paper. A Contextual feature extraction module (CFM) for extracting multi-scale features and an Adaptive Residual Dense Module (ARDN) are used as sub-modules of MSRDNet. Moreover, all the hierarchical features extracted by each ARDN are fused, which helps to detect hazy maps of varying lengths with multi-scale features. This framework outperforms the state-of-the-art dehazing methods in removing haze while maintaining and restoring image detail in real-world and synthetic images captured under various scenarios.
Recovery of true color from underwater images is an ill-posed problem. This is because the wide-band attenuation coefficients for the RGB color channels depend on object range, reflectance, etc. which are difficult to...
详细信息
High-resolution optical images are heavily utilized in various remote sensing applications. The optical images cannot reflect the actual ground information in cloudy conditions. SAR images are used to solve this for t...
详细信息
ISBN:
(纸本)9781450398220
High-resolution optical images are heavily utilized in various remote sensing applications. The optical images cannot reflect the actual ground information in cloudy conditions. SAR images are used to solve this for their ability to see through clouds. But SAR images are usually available with coarser resolutions. So, there is a need to produce an optical image from a SAR image to overcome bad weather and poor resolution in a single go. In this paper, a novel deep learning architecture named EDCGAN is proposed. The proposed architecture is an encoder-decoder-based conditional GAN that uses multi-scale attentive discrimination to get accurate SAR to RGB image translation. In addition, we have used residual connections, spatial & channel-wise attention for better feature representation. A set of extensive experimentations show that this architecture outperforms the existing state-of-the-art method in terms of PSNR, SSIM, and FSIM_c values for the WHU-SENCity dataset.
Communicating with a person having a hearing or speech disability is always a major challenge. Sign Language (SL) is a medium to remove the barrier of such type of communication. It is a very tough task for a common m...
详细信息
Brain Magnetic Resonance Imaging (MRI) is a non-invasive technique that produces high quality images of the brain and is most suitable for analysis and diagnosis. However, these images can be soiled with noise during ...
详细信息
ISBN:
(纸本)9781450398220
Brain Magnetic Resonance Imaging (MRI) is a non-invasive technique that produces high quality images of the brain and is most suitable for analysis and diagnosis. However, these images can be soiled with noise during image acquisition or transmission. The paper is targeted at removing high density salt and pepper noise from such medical images using a denoising technique based on centroidal mean formulation. The presented method is tested on various noisy brain MRI images and the obtained results are promising even for images with high density corruptions, which is suggestive of the resiliency of the algorithm.
Roads in medium-sized indian towns often have lots of traffic but no (or disregarded) traffic stops. This makes it hard for the blind to cross roads safely, because vision is crucial to determine when crossing is safe...
详细信息
ISBN:
(纸本)9781450398220
Roads in medium-sized indian towns often have lots of traffic but no (or disregarded) traffic stops. This makes it hard for the blind to cross roads safely, because vision is crucial to determine when crossing is safe. Automatic and reliable image-based safety classifiers thus have the potential to help the blind to cross indian roads. Yet, we currently lack datasets collected on indian roads from the pedestrian point-of-view, labelled with road crossing safety information. Existing classifiers from other countries are often intended for crossroads, and hence rely on the detection and presence of traffic lights, which is not applicable in indian conditions. We introduce INDRA (indian Dataset for RoAd crossing), the first dataset capturing videos of indian roads from the pedestrian point-of-view. INDRA contains 104 videos comprising of 26k 1080p frames, each annotated with a binary road crossing safety label and vehicle bounding boxes. We train various classifiers to predict road crossing safety on this data, ranging from SVMs to convolutional neural networks (CNNs). The best performing model DilatedRoadCrossNet is a novel single-image architecture tailored for deployment on the Nvidia Jetson Nano. It achieves 79% recall at 90% precision on unseen images. Lastly, we present a wearable road crossing assistant running DilatedRoadCrossNet, which can help the blind cross indian roads in real-time. The project webpage is https://***/Website/***
暂无评论