This paper presents a convolutional neural network (CNN) architecture for segmenting partially transparent shipping bags in RGB-D images of cluttered scenes containing different packaging items in unstructured configurations. The proposed architecture is optimized for training with a limited number of samples with high variability. The analysis of the results with regard to the input type, network architecture, and lighting conditions, proves that including low-resolution depth information improves the segmentation of objects with similar colors and objects in previously unseen lighting conditions, and the high-resolution color photographs greatly improve the segmentation of details. This motivates the proposed multi-input architecture with early feature fusion in order to fully utilize the benefits of high-resolution photographs and low-resolution depth information. The proposed CNN architecture performs successful segmentation of shipping bags in a cluttered environment among packages and items of different colors and materials with irregular shapes. The CNN provides an improvement in accuracy over well-known semantic segmentation architectures while significantly reducing the required processing time, making it suitable for real-time application.

Noise-robustness has become a crucial parameter in Automatic Speech Recognition (ASR) systems today with their increased use in noise-filled real-world environments. One way to address this issue is to develop features that are innately noise-robust. The Kernel Power flow Orientation Coefficients (KPOCs) are a novel feature set based on spectro-temporal analysis that uses a bank of 2D kernels to extract the dominant orientation of the power flow at each point in the auditory spectrogram of the speech signal. The collection of dominant power flow orientation angles forms a novel representation of the speech signal named the Power flow Orientation Spectrogram (POS), which is innately resistant to the spectral masking introduced by the presence of noise and reverberation. This approach not only grants KPOC its noise robustness, but also keeps the number of output coefficients inherently small, thus eliminating the need of the feature dimensionality reduction otherwise necessary in the conventional the spectro-temporal approach. KPOCs performance has been evaluated on three experimental frameworks, and the results have shown that they outperform a number of well-known noise-robust features for average and low SNRs. The relative improvement in Word Recognition Accuracy (WRA) to the classic Mel Frequency Cepstral Coefficients (MFCCs) for the Aurora 2 task goes from 32% up to 190% for SNRs in the range from 10 down to - 5 dB. The experimental results also show that in clean training the performance of KPOC approaches that of the state-of-the-art noise-robust ASR frontends in all noise scenarios for small vocabulary ASR tasks.

In this paper we present a no-reference quality assessment algorithm for highly compressed H.264 videos. By analyzing the spatio-temporal artifacts and their effect on perceived visual quality we produced an array of viable predictors of quality. Using a feature selection method a small sub-set of features, each applied to a specific artefact-domain, was selected for optimal quality estimation. The features are mapped into video quality scores using a simple linear function, which is computationally efficient and minimizes the effect learning of content. The resulting algorithm is evaluated on content independent sets of highly compressed H.264 video sequences from the LIVE video database where it shows high correlation with the subjective scores.

In this paper, we propose a new global motion registration algorithm based on hierarchical motion estimation of a reduced motion vector set, specifically adapted for the properties of the HDR input images, and designed for implementation on handheld devices, demanding low computational complexity and memory load. The procedure is fully automatic and does not require any end-user involvement. The algorithm shows great effectiveness, correctly registering over 95% of the HDR input images captured by free hand.

The focus of this paper is providing a semantic description of images, where each pixel is assigned a label from a pre-defined set of categories. The human visual system is very efficient in recognizing difficult content, making guesses and predictions reinforced by the contextual relationships of neighboring visual stimuli. Motivated by this notion, we approach the image labeling problem by searching for regions with high classification confidence, and use spatial dependencies to improve decisions about the uncertain ones. The higher the number of high confidence predictions, the more relevant information is available and thus a better overall classification can be achieved. Image regions are tested against separate SVM classifiers, and the final labels are computed by applying a post-processing algorithm to the probability outputs. The proposed method is computationally simpler compared to the trending use of graphical models, and experimental results show that it is especially effective in cases of small training sets, achieving results comparable to the state-of-the-art algorithms.

This paper presents an algorithm for automatic detection of the orientation of user generated images. The images can initially be into 3 different orientations. The algorithm utilizes SVM classifier trained over feature vectors of the low-level characteristics of the images in the training set. In order to increase classification accuracy, prior to the SVM classification, the images are hierarchically pre-classified into different groups regarding to the semantic cues they contain, like presence and absence of sky, light, or human faces. Then separate SVM classifier is trained for each group. Also, the paper presents the conclusions of an online survey about the user preferences for software for automatic image orientation detection and gives explanation how those conclusions correspond to the accuracy of the proposed algorithm.

In this paper, an automated algorithm for fusion of differently exposed images is proposed. The algorithm is based on Gaussian/Laplacian pyramid decomposition of the input images. It includes optimization of the pyramid height for best quality results, and decision module for the necessity of the procedure for the recorded scene. The end user involvement in the process of the creation of the output image is completely eliminated, making this algorithm a good choice for use on a mobile platform, as add-on software for low price mobile cameras. Experimental results show high efficiency of the algorithm and excellent visual quality of the resulting images.

In this paper, a selective perceptual-based (SELP) framework is presented to reduce the complexity of popular super-resolution (SR) algorithms while maintaining the desired quality of the enhanced images/video. A perceptual human visual system model is proposed to compute local contrast sensitivity thresholds. The obtained thresholds are used to select which pixels are super-resolved based on the perceived visibility of local edges. Processing only a set of perceptually significant pixels reduces significantly the computational complexity of SR algorithms without losing the achievable visual quality. The proposed SELP framework is integrated into a maximum-a posteriori-based SR algorithm as well as a fast two-stage fusion-restoration SR estimator. Simulation results show a significant reduction on average in computational complexity with comparable signal-to-noise ratio gains and visual quality.

The paper presents a new algorithm for blocking artifacts reduction in low bit rate video. The algorithm addresses the blocking effect on block boundaries, as well as the one inside the blocks. The algorithm works with variety of coding schemes, and does not exploit any information from the coded bitstream. It has extremely low complexity and it is designed to work on a mobile platform in real time. It incorporates two parts: artifacts detection procedure, which decides whether the processing is required for the particular location in the video frame, and artifacts reduction procedure, which performs the actual filtering depending on local image characteristics. The experiments confirm the algorithms effectiveness, as well as its low computational complexity.

In this paper, a selective perceptual-based (SELP) scheme is presented to reduce the complexity of popular super-resolution (SR) algorithms while maintaining the desired quality of the enhanced images/video. A perceptual Human Visual System (HVS) model is proposed to compute the contrast sensitivity threshold for a given background intensity. The obtained thresholds are used to select which pixels are super-resolved based on the perceived visibility of local edges. This is accomplished by estimating the contrast sensitivity threshold locally over a block. Next, the absolute difference between each pixel and its neighbors is computed and compared to the threshold upon which a decision is made to include the pixel in the SR estimator for the next iteration or not. The perceptual model is integrated into a MAP-based SR algorithm as well as a fast ML estimator. Simulation results show up to 47% reduction on average in computational complexity with comparable SNR gains and visual quality.

High compression ratios introduce visible compression artifacts and blurring. To enhance the quality of the compressed video, a super-resolution (SR) scheme is employed. However, the effectiveness of the SR process depends heavily on the amount of subpixel motion that exists between frames in a video sequence, and on the accuracy of the motion vectors (MV). To increase the effectiveness of the SR process, we introduce the concept of added motion prior to video compression. The parameters of this added motion are sent to the receiver with the compressed bit-stream. The proposed SR scheme also exploits existing motion between frames, by estimating MV for each pixel using the original (uncompressed) video. However, since transmitting the pixel-based MV is not practical due to the resulting significant increase in side information, the proposed scheme selects and sends only a very small set of MV based upon their significance in the super-resolution enhancement. The results obtained show a significant improvement in the visual quality of the video sequences, as well as improvement in PSNR.