Identifying similar products in e-commerce is useful in discovering relationships between products, making recommendations, and increasing diversity in search results. Product representation learning is the first step to define a generalized product similarity metric for search. The second step is to extend similarity search to a large scale (e.g., e-commerce catalog scale) without sacrificing quality. In this work, we present a solution that interweaves both steps, i.e., learn representations suited to high quality retrieval using contrastive learning (CL) and retrieve similar items from a large search space using approximate nearest neighbor search (ANNS) to trade-off quality for speed. We propose a CL training strategy for learning uni-modal encoders suited to multi-modal similarity search for e-commerce. We study ANNS retrieval by generating Pareto Frontiers (PFs) without requiring labels. Our CL training strategy doubles retrieval@1 metric across categories (e.g., from 36% to 88% in category C). We also demonstrate that ANNS engine optimization using PFs help select configurations appropriately (e.g., we achieve 6.8× search speed with just 2% drop from the maximum retrieval accuracy in medium size datasets).

Wide-spread availability of broadband internet access and the ubiquity of thin smart end devices such as smart-phones and tablets have led to a trend of moving more services away from the end devices to data centers, commonly referred to as Cloud Computing. For gaming, the stringent network bandwidth requirements of Cloud Gaming and the rapid growth of massively multiplayer online games (MMOG) call for novel content encoding and scene customization schemes to control the bit rate of the streaming video of the game scene. In this paper, a first step has been taken towards this goal by presenting a selective object encoding method to reduce the required network bandwidth and processing power without much impact on the player's quality of experience.

In this letter, a new fast motion estimation (FME) algorithm capable of producing sub-sample motion vectors at low computational-complexity is proposed. Unlike existing FME algorithms, the proposed algorithm considers the low complexity sub-sample accuracy in designing the search pattern for FME. The proposed FME algorithm is designed in such a way that the block distortion measure is modeled as a parametric surface in the vicinity of the integer-sample motion vector; this modeling enables low computational-complexity sub-sample motion estimation (ME) without pixel interpolation. Experimental results on video test sequences show that the proposed FME algorithm reduces computational complexity of integer and sub-sample ME considerably compared with traditional methods at the cost of negligible performance degradation.

A general framework based on histogram equalization for image contrast enhancement is presented. In this framework, contrast enhancement is posed as an optimization problem that minimizes a cost function. Histogram equalization is an effective technique for contrast enhancement. However, a conventional histogram equalization (HE) usually results in excessive contrast enhancement, which in turn gives the processed image an unnatural look and creates visual artifacts. By introducing specifically designed penalty terms, the level of contrast enhancement can be adjusted; noise robustness, white/black stretching and mean-brightness preservation may easily be incorporated into the optimization. Analytic solutions for some of the important criteria are presented. Finally, a low-complexity algorithm for contrast enhancement is presented, and its performance is demonstrated against a recently proposed method.

Due to memory bandwidth limitations and computational complexity considerations in hardware implementations, block matching combined with lscr1error norm and translational motion model is preferred in motion-estimation algorithms. Performance of this scheme is degraded by noise, compression artifacts, rotation, repeating structures, motion boundaries, zooming, and brightness changes. In this work, we present a Bayesian approach to incorporate prior information into block matching. Hypothesis testing is utilized to choose the most applicable prior motion vector and to compute a prior motion-vector distribution and its precision. Prior distribution is then updated with motion-vector likelihood derived from pixel data to obtain the posterior distribution, which is maximized via a search on the feasible motion-vector space.

Local contrast enhancement (LCE) gives a more lively look to an image or video. With a larger difference of a pixel's luma value from its local mean, the eye needs less time to adjust to that local region for a better contrast sensitivity. Since the human eye is trained and conditioned to recognize a natural looking face, the faces in a locally enhanced image may look unnatural even though other parts of the enhanced image is visually more attractive than the original image. To solve this problem, a low-complexity skin-aware local contrast enhancement algorithm is proposed. The proposed skin-aware local contrast enhancement (SALSA) algorithm avoids creating false edges. Experiment results show that SALSA produces natural looking face and non-facial skin regions in the locally enhanced image.

Edges that are visible in color images may not be detected in the corresponding grayscale image. This is due to the neighboring objects having different hues but the same intensities. Hence, a color edge preserving grayscale conversion algorithm is proposed that helps detect color edges using only the luminance component. The algorithm calculates an approximation to the first principal component to form a new set of luminance coefficients instead of using the conventional luminance coefficients. This method can be directly applied to all existing grayscale edge detectors for color edge detection. Processing only one channel instead of three channels results in lower computational complexity compared to other color edge detectors. Experimental results on test images show similar edge detection capabilities to typical color edge detectors at reduced complexity levels.

We present a locally adaptive non-linear (YENI) filter to obtain the unsharp mask of an image. The unsharp mask obtained by the YENI filter preserves the edges in the image while filtering out the local details, which correspond to mid-range frequencies in the spectrum. The enhanced image using this unsharp mask effectively prevents over/under (o/u) shooting artifacts often observed with other unsharp masking techniques. The enhanced frequency range also spans lower frequencies compared to the techniques that are based on Laplacian filter variants. This improves the visual quality of the image, as measured subjectively and objectively in the real-video experiments. Furthermore, since the YENI filter reduces to an IIR filter at each pixel location, it has a low computational complexity.

An unsharp masking type local contrast enhancement method that uses spatially adaptive 2-dimensional (2-D) recursive filters is presented. The proposed 2-D recursive filters are a family of exponential smoothing filters. Three different forms are proposed: separable extension of the 1-dimensional (1-D) case, non-separable filter that uses 4-neighbors, and non-separable filter that uses 8-neighbors. The isotropy of the transfer function improves when moved from the first to the last; hence, better visual quality is obtained in the real video experiments. The computational complexity of the proposed recursive filters, unlike non-recursive filters, is independent of the desired cutoff frequency. This independence enables different levels of smoothing simply by spatially adapting the recursion coefficients without any additional computation

We present a statistical method that uses prediction modeling to decrease the temporally redundant data transmitted back to the sink. The major novelties are fourfold: First, a prediction model is fit to the sensor data. Second, prediction error is utilized to adaptively update the model parameters using hypothesis testing. Third, a data transformation is proposed to bring the sensor sample series closer to weak stationarity. Finally, an efficient implementation is presented. We show that our proposed preDiction eRror bASed hypoThesis testInG (DRASTIG) method achieves low energy dissipation while keeping the prediction errors at user-defined tolerable magnitudes based on real data experiments.

We present adaptive sensing, an energy-efficient topology configuration method for environment monitoring using densely deployed wireless sensor networks. Adaptive sensing puts redundant nodes into passive mode as auxiliary nodes to be used later on, thereby extending the system lifetime. Sensor data is collected at powerful macronodes periodically and a low-order model is fitted to compute a prediction area for each sensor. This prediction area is used to measure the redundancy level of the sensors covered in it. The redundant nodes are put into passive mode by macronodes while keeping the distortion in the output low. We analyze the energy-distortion tradeoff in monitoring applications, introduce the adaptive sensing algorithm based on this analysis. We also include a performance study demonstrating the advantages of our approach based on simulation results.

In this paper, we present PINCO, an in-network compression scheme for energy constrained, distributed, wireless sensor networks. PINCO reduces redundancy in the data collected from sensors, thereby decreasing the wireless communication among the sensor nodes and saving energy. Sensor data is buffered in the network and combined through a pipelined compression scheme into groups of data, while satisfying a user-specified end-to-end latency bound. We introduce a PINCO scheme for single-valued sensor readings. In this scheme, each group of data is a highly flexible structure so that compressed data can be recompressed without decompressing, in order to reduce newly available redundancy at a different stage of the network. We discuss how PINCO parameters affect its performance, and how to tweak them for different performance requirements. We also include a performance study demonstrating the advantages of our approach over other data collection schemes based on simulation and prototype deployment results.