Engagement detection is essential in many areas such as driver attention tracking, employee engagement monitoring, and student engagement evaluation. In this paper, we propose a novel approach using attention based hybrid deep models for the 8th Emotion Recognition in the Wild (EmotiW 2020) Grand Challenge in the category of engagement prediction in the wild EMOTIW2020. The task aims to predict the engagement intensity of subjects in videos, and the subjects are students watching educational videos from Massive Open Online Courses (MOOCs). To complete the task, we propose a hybrid deep model based on multi-rate and multi-instance attention. The novelty of the proposed model can be summarized in three aspects: (a) an attention based Gated Recurrent Unit (GRU) deep network, (b) heuristic multi-rate processing on video based data, and (c) a rigorous and accurate ensemble model. Experimental results on the validation set and test set show that our method makes promising improvements, achieving a competitively low MSE of 0.0541 on the test set, improving on the baseline results by 64%. The proposed model won the first place in the engagement prediction in the wild challenge.

A graph neural network (GNN) for image understanding based on multiple cues is proposed in this paper. Compared to traditional feature and decision fusion approaches that neglect the fact that features can interact and exchange information, the proposed GNN is able to pass information among features extracted from different models. Two image understanding tasks, namely group-level emotion recognition (GER) and event recognition, which are highly semantic and require the interaction of several deep models to synthesize multiple cues, were selected to validate the performance of the proposed method. It is shown through experiments that the proposed method achieves state-of-the-art performance on the selected image understanding tasks. In addition, a new group-level emotion recognition database is introduced and shared in this paper.

Group cohesiveness is a compelling and often studied composition in group dynamics and group performance. The enormous number of web images of groups of people can be used to develop an effective method to detect group cohesiveness. This paper introduces an automatic group cohesiveness prediction method for the 7th Emotion Recognition in the Wild (EmotiW 2019) Grand Challenge in the category of Group-based Cohesion Prediction. The task is to predict the cohesive level for a group of people in images. To tackle this problem, a hybrid network including regression models which are separately trained on face features, skeleton features, and scene features is proposed. Predicted regression values, corresponding to each feature, are fused for the final cohesive intensity. Experimental results demonstrate that the proposed hybrid network is effective and makes promising improvements. A mean squared error (MSE) of 0.444 is achieved on the testing sets which outperforms the baseline MSE of 0.5.

This paper presents a hybrid deep learning network submitted to the 6th Emotion Recognition in the Wild (EmotiW 2018) Grand Challenge [9], in the category of group-level emotion recognition. Advanced deep learning models trained individually on faces, scenes, skeletons and salient regions using visual attention mechanisms are fused to classify the emotion of a group of people in an image as positive, neutral or negative. Experimental results show that the proposed hybrid network achieves 78.98% and 68.08% classification accuracy on the validation and testing sets, respectively. These results outperform the baseline of 64% and 61%, and achieved the first place in the challenge.

We present a rate insensitive method for detecting steganography using plusmn1 embedding. plusmn1 embedding is generally considered to be difficult to detect. In prior work, we developed a high performing method based on statistics computed by a lossless image compressor. However, the method suffered a weakness that the embedding rate needed to be known. In this work, we introduce improved classifiers that work without knowing the rate. The overall classification rate is excellent for the database of images studied.

We present a method for detecting steganography using plusmn1 embedding. The method uses a lossless compression technique to compress the last two bitplanes in an effort to model the image structure where the data may be hidden. A small number of statistics are then computed using the model and fed into a support vector machine to classify detection results. Results presented are obtained using k-fold cross-validation method using a large set of never compressed grayscale images. Detection results show improvement over current plusmn1 steganalysis methods.

We present a new lossless image compression algorithm called BCTW, for bitplane context tree weighting, and a corresponding study into lossless image compression using several number representations and various algorithms. BCTW processes the image bitplane by bitplane and uses context tree weighting (CTW) to estimate the probability of each pixel bit being a 1 or 0. The bit is then compressed with an entropy coder, most likely an arithmetic coder. BCTW is a very good lossless image compressor. In the study, we also look at several number representations (twos-complement, Gray coding, and sign-magnitude) and compare BCTW to MedBzip2 and JPEG-LS. On the Waterloo Bragzone GreySet2 dataset, BCTW outperforms JPEG-LS by 13% and MedBzip2 by 5%. We find that representing the pixels in a Gray code does not help. Sign-magnitude helps sometimes, but not in BCTW. The MED transform helps bzip2, but not BCTW.

In this paper, we present a new approach, the image CRC noise tolerant message authentication code (image CRC-NTMAC), to ensure the credibility of multimedia, such as images. The image CRC-NTMAC can accurately identify the specific regions of the image that have been modified and visibly display the authentication result. Moreover, the tolerance to the low rate random errors, often allowed with multimedia objects, and the sensitivity to the high noise (as forgery) allow the image CRC-NTMAC distinguishing between malicious forgeries and minor communication errors. Examples are provided for the image CRC-NTMAC demonstrating the ability to distinguish between random errors and forgeries. The image CRC-NTMAC improves on the recently developed NTMAC. As with the NTMAC, the image CRC-NTMAC can tolerate a modest number of errors. However, it has thousands of times lower collision rates than the NT-MAC does, thus providing much higher possibility to detect the changes, and tolerating more random channel errors.

This paper introduces a new construct, called the Noise Tolerant Message Authentication Code (NTMAC), for noisy message authentication. The NTMAC can tolerate a small number of errors, such as might be caused by a noisy communications channel. The NTMAC uses a conventional Message Authentication Code (MAC) in its constructions and it inherits the conventional MAC's resistance to forgeries. Furthermore, the NTMAC gives an indication of the number and locations of the errors.

We present a new message authentication code (MAC) called the cyclic redundancy check-noise-tolerant message authentication code (CRC-NTMAC). The CRC-NTMAC improves upon the recently developed NTMAC. Unlike conventional MACs, such as the keyed-hash MAC (HMAC), the CRC-NTMAC and NTMAC tolerate a modest number of errors, such as might be caused by a noisy communications channel. Furthermore, both new MACs provide estimates of the number and locations of the errors. The CRC-NTMAC, however, provides better estimates and has a lower false acceptance probability. The CRC-NTMAC is also not based on cryptographic hash functions

We present a context-weighting algorithm that adaptively weights in real-time three-context models based on their relative accuracy. It can automatically select the better model over different regions of an image, producing better probability estimates than using either one of these models exclusively. Combined with the previously proposed block arithmetic coder for image compression (BACIC), the overall performance is slightly better than JBIG for the eight CCITT business-type test images, outperforms JBIG by 13.8% on halftone images, and by 17.5% for compounded images containing both text and halftones. Furthermore, users no longer need to select models as in JBIG and BACIC to get the better performance.

We present a new method for authenticating and tamperproofing images. It is based on the recently developed NT-MAC (noise tolerant message authentication code). The NTMAC is designed for multimedia transmitted over noisy channels. Unlike ordinary cryptographic codes, the NT-MAC fails gradually in the presence of noise or bit errors. In this work, we show how a combination of structural and random partitions allows the NTMAC to identify regions which might represent forgeries while distinguishing between forgeries and minor errors caused by communications noise. The paper concludes with a simple example of a forged image and the NTMAC's estimate of the changes.

A context-weighting algorithm as an improvement of previously proposed block arithmetic coder for image compression (BACIC) is presented. The proposed algorithm weights two context models, so that it can automatically select the better model over different regions of an image, producing better probability estimates. The overall performance of this algorithm is better than single context BACIC; it is the same as JBIG1 for the eight CCITT business-type test images, and outclasses JBIG1 by 13.8% on halftone images, by 25.7% for images containing both text and halftones. Furthermore, users no longer need to select models as in JBIG1 and BACIC to get the better performance. Rather than using segmentation algorithm in JBIG2, the context-weighting BACIC may be a good choice in many applications due to its simplicity.

We present a two-part method for lossless image compression: the first is a simple transformation based on JPEG-LS and the second is the use of a general purpose file compressor such as gzip, bzip2 or rk. The transformation is invertible and uses only integer arithmetic. It achieves no compression directly; however, on some images, it can significantly reduce the entropy. The gzip, bzip2, and rk compressors, with and without the transformation, are tested against JPEG-LS on a standard data set. Overall rk (both with and without the transformation) is best. Furthermore, by combining the various algorithms, we obtain compression that is 15% better than JPEG-LS alone.

This paper presents the block arithmetic coding for image compression (BACIC) algorithm: a new method for lossless bilevel image compression which can replace JBIG, the current standard for bilevel image compression. BACIC uses the block arithmetic coder (BAC): a simple, efficient, easy-to-implement, variable-to-fixed arithmetic coder, to encode images. BACIC models its probability estimates adaptively based on a 12-bit context of previous pixel values; the 12-bit context serves as an index into a probability table whose entries are used to compute p(1) (the probability of a bit equaling one), the probability measure BAC needs to compute a codeword. In contrast, the Joint Bilevel Image Experts Group (JBIG) uses a patented arithmetic coder, the IBM QM-coder, to compress image data and a predetermined probability table to estimate its probability measures. JBIG, though, has not get been commercially implemented; instead, JBIG's predecessor, the Group 3 fax (G3), continues to be used. BACIC achieves compression ratios comparable to JBIG's and is introduced as an alternative to the JBIG and G3 algorithms. BACIC's overall compression ratio is 19.0 for the eight CCITT test images (compared to JBIG's 19.6 and G3's 7.7), is 16.0 for 20 additional business-type documents (compared to JBIG's 16.0 and G3's 6.74), and is 3.07 for halftone images (compared to JBIG's 2.75 and G3's 0.50).

This study introduces error control to the block arithmetic coding for image compression (BACIC): a new method for lossless bilevel image compression. BACIC can successfully transmit bilevel images when channel bit error rates are as high as 10(-3) while providing compression ratios twice that of G3, the only facsimile standard which incorporates error control into its algorithm.

In this paper, we present a new method of digital steganography, entitled spread spectrum image steganography (SSIS). Steganography, which means "covered writing" in Greek, is the science of communicating in a hidden manner. Following a discussion of steganographic communication theory and review of existing techniques, the new method, SSIS, is introduced. This system hides and recovers a message of substantial length within digital imagery while maintaining the original image size and dynamic range. The hidden message can be recovered using appropriate keys without any knowledge of the original image. Image restoration, error-control coding, and techniques similar to spread spectrum are described, and the performance of the system is illustrated. A message embedded by this method can be in the form of text, imagery, or any other digital signal. Applications for such a data-hiding scheme include in-band captioning, covert communication, image tamperproofing, authentication, embedded control, and revision tracking.

We present a new method of embedding information within digital images, called spread spectrum image steganography (SSIS). Steganography, which means "covered writing" in Greek, is the science of communicating in a hidden manner. SSIS conceals a message of substantial length within digital imagery while maintaining the original image size and dynamic range. The hidden message can be recovered using the appropriate keys without any knowledge of the original image. Image processing, error control coding, and spread spectrum techniques used to conceal the hidden data are described, and the performance of the technique is illustrated. The message embedded by this method can be in the form of text, imagery, or any other digital signal. Applications for such a data-hiding scheme include in-band captioning, hidden communication, image tamperproofing, authentication, invisible map overlays, embedded control, and revision tracking.

BACIC is a new method of lossless bi-level image compression introduced to replace JBIG and G3, the current standards for bi-level and facsimile image compression. This paper applies the BACIC (block arithmetic coding for image compression) algorithm to reduced grayscale and full grayscale image compression. BACIC's compressed files are slightly smaller than JBIGs, twice as small as G3s, and at least thirty percent smaller than lossless-JPEGs (when lossless JPEG uses Huffman coding) for reduced grayscale images with fewer than 7 bits/pixel.

We introduce a new multiresolution watermarking method for digital images. The method is based on the discrete wavelet transform (DWT). Pseudo-random codes are added to the large coefficients at the high and middle frequency bands of the DWT of an image. It is shown that this method is more robust to often proposed methods to some common image distortions, such as the wavelet transform based image compression, and image halftoning. Moreover, the method is hierarchical. The computation load needed to detect the watermark depends on the noise level in an image.

Robust source coding provides both compression and noise mitigation without channel coding. Two methods of robust source coding are presented. The first is DPCM incorporating a nonlinear filter; the second is predictive trellis-coded quantization (PTCQ), whose prediction filter is also nonlinear. Findings show that the incorporation of the nonlinear filter provides the ability to conceal the effects of noise and minimize the propagation of error. A comparison is made between the performance of the two methods in a noisy environment. It is demonstrated that the PTCQ scheme provides a method of least distortion at low BER. However, at higher BER the DPCM scheme outperforms PTCQ. Robust source coding of images employing PTCQ or DPCM with the nonlinear prediction filter exhibits low sensitivity to bit errors and yields, visually, a better reconstructed image after transmission over very noisy channels.

Finding optimal kinematics for manipulators dedicated to working within enclosed workspaces can be applied to assembly-line tasks such as welding, grinding, and spray-painting within automobile interiors. This paper presents a method of designing redundant and hyper-redundant manipulators for enclosed workspaces. A heuristic search is used to generate designs capable of maneuvering within a specified target-workspace in a given enclosure. A locally-optimal design is then selected for minimum degrees of freedom, or for performance values measured during the search. A variation on task-priority based redundancy control is also presented which allows many secondary-priority obstacle avoidance tasks to be satisfied simultaneously, therefore permitting easier maneuvering through complex enclosures.

This correspondence examines the use of nonlinear edge enhancers as prefilters for edge detectors. The filters are able to convert smooth edges to step edges and suppress noise simultaneously. Thus, false alarms due to noise are minimized and edge gradient estimates tend to be large and localized. This leads to significantly improved edge maps.

We introduce "Block Arithmetic Coding" (BAC), a technique for entropy coding that combines many of the advantages of ordinary stream arithmetic coding with the simplicity of block codes. The code is variable length in to fixed out (V to F), unlike Huffman coding which is fixed in to variable out (F to V). We develop two versions of the coder: 1) an optimal encoder based on dynamic programming arguments, and 2) a suboptimal heuristic based on arithmetic coding. The optimal coder is optimal over all V to F complete and proper block codes. We show that the suboptimal coder achieves compression that is within a constant of a perfect entropy coder for independent and identically distributed inputs. BAC is easily implemented, even with large codebooks, because the algorithms for coding and decoding are regular. For instance, codebooks with 2/sup 32/ entries are feasible. BAC also does not suffer catastrophic failure in the presence of channel errors. Decoding errors are confined to the block in question. The encoding is in practice reasonably efficient. With i.i.d. binary inputs with P(1)=0.95 and 16 bit codes, entropy arguments indicate at most 55.8 bits can be encoded; the BAC heuristic achieves 53.0 and the optimal BAC achieves 53.5. Finally, BAC appears to be much faster than ordinary arithmetic coding.>

This paper describes a scheduling algorithm for high speed, TDMA communication networks. The problem is to find a route and determine switch schedules so that a burst of information can traverse the network from source to destination. We show that this problem is difficult, i.e., that several simple variations are NP-complete. We present an algorithm that finds schedules, and within a simplifying assumption, is guaranteed to find a route if one exists. Simulations indicate that this algorithm runs quickly in practice and overall scheduling efficiencies of 80-90% can be achieved. We also describe a variation in which the problem is to find a repeating schedule and we show how to adapt the basic algorithm for this variation.

Algorithms for computing the distributions of order statistic related estimators with moving or multiple windows are presented. These algorithms may be used to compute joint distributions of moving window estimators, such as moving median filters, or of estimators made from ranking operations on multiple windows, such as many stacked or morphological filters. The presented algorithms make no distributional assumptions on the underlying random variables, but do make assumptions on the dependency between them. For instance, the random variables may be independent, Markov, or Markov-corrupted by an independent noise source. Unlike other approaches, these algorithms have polynomial complexity in the number of random variables. How these algorithms may be easily implemented is shown. Finally, two computational examples of the behavior of median filters are given.>

A generalized framework for the description of a large class of nonlinear filters is introduced. This framework includes nonrecursive linear filters (FIR), order statistic filters (OSF or L-filters), Ll-filters, Volterra filters, and ZNL-LTI filters. The Ll-filters have been proposed to generalize and combine FIR and OSF filters. LMS and RLS algorithms are extended to such a class, and a simulation of an adaptive line enhancer using an L-filter is presented.>