-
Two Heads are Better Than One: Neural Networks Quantization with 2D Hilbert Curve-based Output Representation
Authors:
Mykhailo Uss,
Ruslan Yermolenko,
Olena Kolodiazhna,
Oleksii Shashko,
Ivan Safonov,
Volodymyr Savin,
Yoonjae Yeo,
Seowon Ji,
Jaeyun Jeong
Abstract:
Quantization is widely used to increase deep neural networks' (DNN) memory, computation, and power efficiency. Various techniques, such as post-training quantization and quantization-aware training, have been proposed to improve quantization quality. We introduce a novel approach for DNN quantization that uses a redundant representation of DNN's output. We represent the target quantity as a point…
▽ More
Quantization is widely used to increase deep neural networks' (DNN) memory, computation, and power efficiency. Various techniques, such as post-training quantization and quantization-aware training, have been proposed to improve quantization quality. We introduce a novel approach for DNN quantization that uses a redundant representation of DNN's output. We represent the target quantity as a point on a 2D parametric curve. The DNN model is modified to predict 2D points that are mapped back to the target quantity at a post-processing stage. We demonstrate that this map** can reduce quantization error. For the low-order parametric Hilbert curve, Depth-From-Stereo task, and two models represented by U-Net architecture and vision transformer, we achieved a quantization error reduction by about 5 times for the INT8 model at both CPU and DSP delegates. This gain comes with a minimal inference time increase (less than 7%). Our approach can be applied to other tasks, including segmentation, object detection, and key-points prediction.
△ Less
Submitted 22 May, 2024;
originally announced May 2024.
-
Estimation of Variance and Spatial Correlation Width for Fine-scale Measurement Error in Digital Elevation Model
Authors:
Mykhail Uss,
Benoit Vozel,
Vladimir Lukin,
Kacem Chehdi
Abstract:
In this paper, we borrow from blind noise parameter estimation (BNPE) methodology early developed in the image processing field an original and innovative no-reference approach to estimate Digital Elevation Model (DEM) vertical error parameters without resorting to a reference DEM. The challenges associated with the proposed approach related to the physical nature of the error and its multifactor…
▽ More
In this paper, we borrow from blind noise parameter estimation (BNPE) methodology early developed in the image processing field an original and innovative no-reference approach to estimate Digital Elevation Model (DEM) vertical error parameters without resorting to a reference DEM. The challenges associated with the proposed approach related to the physical nature of the error and its multifactor structure in DEM are discussed in detail. A suitable multivariate method is then developed for estimating the error in gridded DEM. It is built on a recently proposed vectorial BNPE method for estimating spatially correlated noise using Noise Informative areas and Fractal Brownian Motion. The newly multivariate method is derived to estimate the effect of the stacking procedure and that of the epipolar line error on local (fine-scale) standard deviation and autocorrelation function width of photogrammetric DEM measurement error. Applying the new estimator to ASTER GDEM2 and ALOS World 3D DEMs, good agreement of derived estimates with results available in the literature is evidenced. In future works, the proposed no-reference method for analyzing DEM error can be extended to a larger number of predictors for accounting for other factors influencing remote sensing (RS) DEM accuracy.
△ Less
Submitted 23 January, 2018;
originally announced January 2018.
-
Multimodal Remote Sensing Image Registration with Accuracy Estimation at Local and Global Scales
Authors:
M. L. Uss,
B. Vozel,
V. V. Lukin,
K. Chehdi
Abstract:
This paper focuses on potential accuracy of remote sensing images registration. We investigate how this accuracy can be estimated without ground truth available and used to improve registration quality of mono- and multi-modal pair of images. At the local scale of image fragments, the Cramer-Rao lower bound (CRLB) on registration error is estimated for each local correspondence between coarsely re…
▽ More
This paper focuses on potential accuracy of remote sensing images registration. We investigate how this accuracy can be estimated without ground truth available and used to improve registration quality of mono- and multi-modal pair of images. At the local scale of image fragments, the Cramer-Rao lower bound (CRLB) on registration error is estimated for each local correspondence between coarsely registered pair of images. This CRLB is defined by local image texture and noise properties. Opposite to the standard approach, where registration accuracy is only evaluated at the output of the registration process, such valuable information is used by us as an additional input knowledge. It greatly helps detecting and discarding outliers and refining the estimation of geometrical transformation model parameters. Based on these ideas, a new area-based registration method called RAE (Registration with Accuracy Estimation) is proposed. In addition to its ability to automatically register very complex multimodal image pairs with high accuracy, the RAE method provides registration accuracy at the global scale as covariance matrix of estimation error of geometrical transformation model parameters or as point-wise registration Standard Deviation. This accuracy does not depend on any ground truth availability and characterizes each pair of registered images individually. Thus, the RAE method can identify image areas for which a predefined registration accuracy is guaranteed. The RAE method is proved successful with reaching subpixel accuracy while registering eight complex mono/multimodal and multitemporal image pairs including optical to optical, optical to radar, optical to Digital Elevation Model (DEM) images and DEM to radar cases. Other methods employed in comparisons fail to provide in a stable manner accurate results on the same test cases.
△ Less
Submitted 25 May, 2016; v1 submitted 8 February, 2016;
originally announced February 2016.
-
Efficient Rotation-Scaling-Translation Parameters Estimation Based on Fractal Image Model
Authors:
M. Uss,
B. Vozel,
V. Lukin,
K. Chehdi
Abstract:
This paper deals with area-based subpixel image registration under rotation-isometric scaling-translation transformation hypothesis. Our approach is based on a parametrical modeling of geometrically transformed textural image fragments and maximum likelihood estimation of transformation vector between them. Due to the parametrical approach based on the fractional Brownian motion modeling of the lo…
▽ More
This paper deals with area-based subpixel image registration under rotation-isometric scaling-translation transformation hypothesis. Our approach is based on a parametrical modeling of geometrically transformed textural image fragments and maximum likelihood estimation of transformation vector between them. Due to the parametrical approach based on the fractional Brownian motion modeling of the local fragments texture, the proposed estimator MLfBm (ML stands for "Maximum Likelihood" and fBm for "Fractal Brownian motion") has the ability to better adapt to real image texture content compared to other methods relying on universal similarity measures like mutual information or normalized correlation. The main benefits are observed when assumptions underlying the fBm model are fully satisfied, e.g. for isotropic normally distributed textures with stationary increments. Experiments on both simulated and real images and for high and weak correlation between registered images show that the MLfBm estimator offers significant improvement compared to other state-of-the-art methods. It reduces translation vector, rotation angle and scaling factor estimation errors by a factor of about 1.75...2 and it decreases probability of false match by up to 5 times. Besides, an accurate confidence interval for MLfBm estimates can be obtained from the Cramer-Rao lower bound on rotation-scaling-translation parameters estimation error. This bound depends on texture roughness, noise level in reference and template images, correlation between these images and geometrical transformation parameters.
△ Less
Submitted 4 July, 2015; v1 submitted 10 January, 2015;
originally announced January 2015.
