The latest research in computer vision highlighted the effectiveness of the vision transformers (ViT) in performing several computer vision tasks; they can efficiently understand and process the image globally unlike the convolution which processes the image locally. ViTs outperform the convolutional neural networks in terms of accuracy in many computer vision tasks but the speed of ViTs is still an issue, due to the excessive use of the transformer layers that include many fully connected layers. Therefore, we propose a real-time ViT-based monocular depth estimation (depth estimation from single RGB image) method with encoder-decoder architectures for indoor and outdoor scenes. This main architecture of the proposed method consists of a vision transformer encoder and a convolutional neural network decoder. We started by training the base vision transformer (ViT-b16) with 12 transformer layers then we reduced the transformer layers to six layers, namely ViT-s16 (the Small ViT) and four layers, namely ViT-t16 (the Tiny ViT) to obtain real-time processing. We also try four different configurations of the CNN decoder network. The proposed architectures can learn the task of depth estimation efficiently and can produce more accurate depth predictions than the fully convolutional-based methods taking advantage of the multi-head self-attention module. We train the proposed encoder-decoder architecture end-to-end on the challenging NYU-depthV2 and CITYSCAPES benchmarks then we evaluate the trained models on the validation and test sets of the same benchmarks showing that it outperforms many state-of-the-art methods on depth estimation while performing the task in real-time (∼20 fps). We also present a fast 3D reconstruction (∼17 fps) experiment based on the depth estimated from our method which is considered a real-world application of our method.

We propose a high-speed convolutional neural network approach for weakly supervised localization (WSL) and weakly supervised object detection (WSOD). The proposed method, called center-of-features localization (COFL), performs localization of objects in a visual scene by combining both multi-label classification and regression for the number of instances of each class object. A modified Xception network architecture is used as the main feature extractor, and a classification-plus-regression loss function is used to perform the detection task. The method does not require bounding box annotations but only image labels and counts of the objects of each class in the image. This combination can produce a clear localization of objects in the scene through a masking technique between class activation maps (CAMs) and regression activation maps (RAMs). The proposed method was trained and tested on the PASCAL VOC2007 and VOC2012 datasets; it attained a mean average precision (mAP) of 47.0% and a correct localization CorLoc of 64.1% on PASCAL VOC2007 and a mAP of 42.3% and a CorLoc of 65.5% on PASCAL VOC2012 while performing object detection at a speed of ~50 fps. These results demonstrate that the network can perform object detection accurately in real-time using only image labels and object counts, which are inexpensive to annotate compared with the bounding box annotations typically employed in fully supervised object detection methods. The network far outperforms other weakly supervised methods and some fully supervised methods in terms of processing time while achieving fair accuracy.

Although Light-Field (LF) technology attracts attention due to its large number of applications, especially with the introduction of consumer LF cameras and its frequent use, reconstructing densely sampled LF images represents a great challenge to the use and development of LF technology. Our paper proposes a learning-based method to reconstruct densely sampled LF images from a sparse set of input images. We trained our model with raw LF images rather than using multiple images of the same scene. Raw LF can represent the two-dimensional array of images captured in a single image. Therefore, it enables the network to understand and model the relationship between different images of the same scene well and thus restore more texture details and provide better quality. Using raw images has transformed the task from image reconstruction into image-to-image translation. The feature of small-baseline LF was used to define the images to be reconstructed using the nearest input view to initialize input images. Our network was trained end-to-end to minimize the sum of absolute errors between the reconstructed and ground-truth images. Experimental results on three challenging real-world datasets demonstrate the high performance of our proposed method and its outperformance over the state-of-the-art methods.

Dense multi-view image reconstruction has been a focal point of research for an extended period, with recent surges in interest. The utilization of multi-view images offers solutions to numerous challenges and amplifies the effectiveness of various applications including 3D reconstruction, de-occlusion, depth sensing, saliency detection, and identifying salient objects. This paper introduces an approach to reconstructing high-density light field (LF) images, addressing the inherent challenge of balancing angular and spatial resolution caused by limited sensor resolution. We introduce an innovative approach to reconstructing LF images through a CNN-based network that combines spatial and epipolar features in both initial and deep feature extraction phases. Our network utilizes angular information during upsampling and employs dual feature extraction to effectively analyze horizontal and vertical epipolar data. Weight sharing within the CNN block between horizontal and vertically transposed stacks enhances quality while preserving model compactness. The outcomes of experiments carried out on real-world and synthetic datasets demonstrate the effectiveness of our method, showcasing its superior performance in both inference speed and reconstruction quality when compared to state-of-the-art (SOTA) techniques.

Chest X-ray (CXR) imaging plays a pivotal role in diagnosing various pulmonary diseases, which account for a significant portion of the global mortality rate, as recognized by the World Health Organization (WHO). Medical practitioners routinely depend on CXR images to identify anomalies and make critical clinical decisions. Dramatic improvements in super-resolution (SR) have been achieved by applying deep learning techniques. However, some SR methods are very difficult to utilize due to their low-resolution inputs and features containing abundant low-frequency information, similar to the case of X-ray image super-resolution. In this paper, we introduce an advanced deep learning-based SR approach that incorporates the innovative residual-in-residual (RIR) structure to augment the diagnostic potential of CXR imaging. Specifically, we propose forming a light network consisting of residual groups built by residual blocks, with multiple skip connections to facilitate the efficient bypassing of abundant low-frequency information through multiple skip connections. This approach allows the main network to concentrate on learning high-frequency information. In addition, we adopted the dense feature fusion within residual groups and designed high parallel residual blocks for better feature extraction. Our proposed methods exhibit superior performance compared to existing state-of-the-art (SOTA) SR methods, delivering enhanced accuracy and notable visual improvements, as evidenced by our results.

We propose Depth-to-Space Net (DTS-Net), an effective technique for semantic segmentation using the efficient sub-pixel convolutional neural network. This technique is inspired by depth-to-space (DTS) image reconstruction, which was originally used for image and video super-resolution tasks, combined with a mask enhancement filtration technique based on multi-label classification, namely, Nearest Label Filtration. In the proposed technique, we employ depth-wise separable convolution-based architectures. We propose both a deep network, that is, DTS-Net, and a lightweight network, DTS-Net-Lite, for real-time semantic segmentation; these networks employ Xception and MobileNetV2 architectures as the feature extractors, respectively. In addition, we explore the joint semantic segmentation and depth estimation task and demonstrate that the proposed technique can efficiently perform both tasks simultaneously, outperforming state-of-art (SOTA) methods. We train and evaluate the performance of the proposed method on the PASCAL VOC2012, NYUV2, and CITYSCAPES benchmarks. Hence, we obtain high mean intersection over union (mIOU) and mean pixel accuracy (Pix.acc.) values using simple and lightweight convolutional neural network architectures of the developed networks. Notably, the proposed method outperforms SOTA methods that depend on encoder–decoder architectures, although our implementation and computations are far simpler.