Stereo image Super-Resolution (SR) aims to enhance image resolution by leveraging complementary information in stereo pairs. Convolutional Neural Networks (CNNs), widely used in stereo image SR for their strong local pattern extraction capabilities, often fail to capture long-range dependencies critical for stereo correspondence. On the other hand, Swin Transformers have demonstrated superior performance in modeling long-range dependencies for stereo image SR tasks. However, their computational complexity scales quadratically with the window size, leading to a trade-off between global receptive fields and computational efficiency. To tackle these challenges, we propose StereoMamba+, a novel stereo image SR method designed to adaptively capture both local and global dependencies in stereo pairs. Leveraging the Mamba architecture as its backbone, StereoMamba+ integrates an Adaptive State Space Module (ASSM) that efficiently extracts and fuses global and local features, maintaining linear computational complexity. Additionally, a Gated Enhanced Feed-Forward Network (GEFN) selectively amplifies essential features and depth cues, and a Residual Frequency Block (RFB) is employed to capture global features in the frequency domain. To further enhance stereo correspondence, we introduce a Stereo Bi-Directional Cross Attention Module (SBCAM), aligning unique features along both horizontal and vertical epipolar lines to improve stereo consistency. Extensive experiments demonstrate that our proposed StereoMamba+ method achieves state-of-the-art performance on 2× and 4× stereo image SR tasks, delivering PSNR improvements of up to 0.45dB, while maintaining competitive parameter efficiency compared to existing methods.

Stereo image super-resolution (SR) aims to enhance image resolution by leveraging the complementary information from stereo image pairs. While convolutional neural network (CNN)-based methods have traditionally dominated this field, they struggle with capturing long-range dependencies. Transformer-based approaches have shown improvements by better modeling long-range dependencies, but their computational complexity scales quadratically with respect to the window length. To address these challenges, in this paper we propose StereoMamba, a new stereo image super-resolution method built on Structured State Space Models (SSMs). StereoMamba leverages the Mamba architecture to effectively capture long-range dependencies and inter-view correlations in stereo image pairs. Additionally, we introduce a Stereo Bi-directional Cross-Attention Module (SBCAM) to further improve stereo view correlation. Extensive experiments show that StereoMamba consistently surpasses state-of-the-art methods across several public datasets.

The use of Stable Diffusion models to generate realistic images has become a popular topic in recent years. However, this technology has also raised concerns about the potential harm it may cause to the copyright holders, particularly in the realm of art where these synthesized images can closely resemble the original work. As these synthesized images are hard for humans to distinguish from authentic ones, it is of great importance to develop methods that may identify them. In this paper, we propose a deep learning-based approach to detect synthesized images using information in the frequency domain. Since there exists no well-established dataset of images synthesized by stable diffusion models, in order to train and evaluate our network we generated a representative dataset consisting of carefully selecting human-created authentic images and synthesized animation images generated by the Stable Diffusion models. We chose to use Disney-style animated content for our case study, given its significance in the realm of intellectual property protection. Experimental results demonstrated that our proposed model outperforms humans and other state-of-the-art methods, achieving an accuracy rate of 99.46%.

In this paper, a new and calibrated light field (LF) video dataset is introduced, which focuses on outdoor scenes and objects. Each video stream is 10 seconds long and it is captured with a dense camera array that consists of$5\times 5$camera modules in$1640\times 1232$resolution at 40 frames per second. As multiple cameras in an array setup may suffer from various conditions of camera settings, lens structure, and lighting variations, the resulting images can be negatively affected by geometric distortion and colour difference. To address that, a unified calibration method involving both spatial calibration and colour correction is employed to correct inconsistences and achieve a better image quality with reduced image distortion. This video dataset would be suitable for further research and investigation of a variety LF applications, such as autonomous driving and immersive media.

Impressive advancements in capturing, display, and broadcasting technologies significantly elevate image and video quality, and with that the need for designing new reference and no-reference image and video quality metrics. One of the latest and perceptually accurate video quality metrics is the Video Multi-Method Assessment Fusion (VMAF) method. However, VMAF considers the temporal nature of video using basic average temporal pooling, an approach that falls short from human perception. In this paper, we introduce a new no-reference video quality metric that uses deep learning to extract spatial features and a unique temporal pooling approach to accurately predict the visual quality score. To this end, first we created a video quality dataset that consists of high-resolution, 20s-long test video clips compressed at several different bitrates. These videos were labeled based on subjective evaluations and were used to determine the perceptual importance of frames in our temporal pooling scheme. Evaluations showed that our proposed approach achieved correlation of 90.55% with human perception and outperformed the state-of-the-art VMAF approach by 15.63% accuracy.

Timely road condition inspection and maintenance are key components of infrastructure management for smart cities, as they reduce traffic congestion, accidents and repairing costs. Traditional road inspection methods that employ vibrations and/or laser scanning for detecting road deterioration use expensive equipment and dedicated municipality vehicles. Recently, computer vision techniques and artificial intelligence are emerging as alternative solutions to traditional approaches for road condition detection, offering more flexibility, higher accuracy, and overall lower cost. In this paper, we utilize convolutional neural network-based and vision transformer-based object detection models to accurately identify road deteriorations namely, potholes, cracks, and alligators. We compare four different state-of-the-art models in terms of detection accuracy and speed. Performance evaluations have shown that, on the same dataset the Swin Transformer model outperformed the other state-of-the-art methods by a substantial margin. With 74% detection accuracy, and 42 frames per second processing speed Swin Transformer exceled over EfficentDet, YOLOv4, and YOLOX. We also present a new comprehensive and balanced large-scale road condition dataset of 27,298 annotated images, captured by ordinary car cameras.

Automatic human real-time fall detection is a challenging task in remote healthcare, demanding a non-intrusive, secure and affordable solution. In this paper, we present a real-time hardware system that uses a deep learning model for fall detection embedded in a color camera. To reduce the startup delay and achieve real-time performance for the inference phase, we optimized our model using TensorRT. In addition, we addressed the board memory limitation using virtual memory and linear memory allocation and garbage collection. Moreover, GStreamer was used to perform most of the video processing using Jetson's GPU. Our live evaluation shows that our system achieved the accuracy of 84.44% and real-time performance.

SHVC, the scalable extension of High Efficiency Video Coding (HEVC), uses advanced inter-layer prediction features in addition to the advanced compression tools of HEVC to improve the compression performance. Using combined features has brought us improved compression performance at the cost of huge computational complexity for the SHVC encoder. This complexity is mainly because of the the inter/intra-prediction mode search of the coding units. The focus of this study is on developing an efficient complexity reduction for quality scalability of SHVC encoder, with the intention to facilitate the adoption of SHVC for real-time applications. In this regard, first, we build a probabilistic model that uses the mode information and motion homogeneity of already encoded blocks in the enhancement layer (EL) and the base layer to predict the probabilities of all the available inter/intra modes of the to-be-coded block in the EL. Then, we propose an online-learning-based fast mode, assigning (FMA) method that uses the proposed probabilistic model to predict the mode of the to-be-coded block in the EL. Performance evaluation shows that our proposed FMA method reduces the total execution time of the SHVC encoder by 45.40% on average compared with unmodified SHVC codec while maintaining the overall video quality.

3-D High Efficiency Video Coding (HEVC) is a new emerging video compression standard for multiview video applications. This standard utilizes advanced interview prediction characteristics in addition to the prediction features of the HEVC standard for efficient encoding of multiview video content. While using combined features improves the compression performance by utilizing the correlation between the views captured from slightly different angles of the same scene, they also increase coding complexity. The focus of this paper is on developing an efficient complexity reduction scheme for 3D-HEVC, with the intention to facilitate the adoption of this upcoming standard, especially for real-time applications. In this regard, first, we introduce two ways to decrease the complexity of the inter-/ intra-mode search process of the to-be-encoded blocks in the dependent texture views (${\mathrm {DV}}_{t}\text{s}$ ) of 3D-HEVC. Then, we propose a hybrid complexity reduction scheme that utilizes the two-mode prediction approaches, motion information of the base texture view (BVt), and the rate distortion cost of the already encoded blocks in the BVt and DVt. The performance of our proposed scheme is tested for the case with two views (i.e., base view + dependent view). The evaluations confirm that our proposed hybrid complexity reduction scheme reduces the 3D-HEVC codec complexity by 67.70% on average for the DVt compared with the unmodified 3D-HEVC encoder, while maintaining the overall video quality. Compared with the state-of-the-art method, it reduces complexity by 25.74% on average.

The scalable extension of HEVC (known as SHVC), the recent scalable video coding standard, results in an improved compression performance at the cost of significant increase in computational coding complexity. One of the main factors that contribute to the SHVC encoder complexity is choosing the best partitioning structure for the coding tree units (CTUs). Our study focuses on developing a scheme for predicting the CTU structure in the quality and spatial scalable extension of HEVC. The proposed scheme uses the CTU partitioning structure of the already encoded CTUs in the enhancement layers (ELs) and base layer (BL) to predict the coding unit sizes of the to-be-encoded CTUs in the EL. Performance evaluations confirm that our proposed complexity reduction scheme significantly reduces the execution time of the SHVC encoder, while maintaining the overall quality of the coded streams.

The 3D extension of High Efficiency Video Coding (HEVC) standard (3D-HEVC) aims at improving coding efficiency by introducing new and unique approaches for utilizing correlations between the different views of a scene. Reported coding efficiency, however, comes at the expense of increased computational complexity. For real-time applications, reducing the computational complexity of 3D-HEVC is very important. In this paper, we propose an adaptive fast mode assigning method based on a Bayesian classifier that reduces 3D-HEVC's coding complexity by up to 51.95%, while maintaining the overall quality and bitrate.

This paper presents a novel quantization-based information hiding approach, which is invariant to gain attack. In the presented scheme, the host signal vector is first divided into two parts, then lp-norm of each vector is calculated. The watermark bits are embedded by quantizing the ratio of the lp-norm of each part. The decoding scheme is performed blindly using the Euclidean distance. Due to the use of division function, the proposed method is robust to scaling attacks. The analytical error probability and embedding distortion are derived and assessed by simulations on artificial signals. The optimum parameter in the lp-norm function is obtained based on minimizing the error probability. Experimental results confirm the superiority of the proposed technique against common attacks in comparison with the existing state-of-the-art methods.

There has been significant interest in developing a scalable version of the High Efficiency Video Coding (HEVC) standard. As expected, the HEVC scalable video version increases the complexity of the codec compared to the non-scalable counterpart. In this paper, we propose an adaptive early-termination interlayer motion prediction mode search that significantly reduces HEVC/SVC's coding complexity by up to 85.77%, while maintaining the overall bitrate.

In this paper, a novel quantization-based information hiding approach which is invariant to gain attack is presented. For the data embedding, the host vector signal is first divided into two separate vectors. Then the ratio of the magnitude of the vectors is quantized according to the watermark data. The decoding scheme is performed blindly using the euclidean distance. The performance of the proposed method is analytically studied and assessed by simulations on artificial signals. The proposed method is applied to various test images as well. The experimental results confirm the superiority of the proposed technique against common attacks in comparison with the recently proposed methods.

In this study, a robust image watermarking based on the quantisation index modulation (QIM) method is proposed. Conventional QIM methods employ a fixed quantisation step‐size that results in poor robustness of the algorithm. Here, the quantisation step‐size in the QIM method is adaptively selected using a power‐law function and with the aid of the side information, the proposed method is invariant to gain and rotation attack. To keep the watermark imperceptible and increase its robustness, the low‐frequency components of high‐entropy image blocks are used for data hiding. The analytical error probability and embedding distortion are derived and assessed by simulations on artificial signals. The optimum parameter in the power‐law function is obtained based on minimising the error probability. Experimental results confirm the superiority of the proposed technique against common attacks in comparison with the recently proposed methods.