A Perceptually Optimized Embedded Image Coder and Quality Assessor Based Both on Visual Tools

Abstract

In this paper, we propose a new Perceptually Optimized Embedded Zero Tree Image Coder EVIC and its improved version MEVIC. The coder presents a new perceptual model to weight wavelet coefficients to enhance SPIHT embedded coding. The new visual coder aims to improve the visual quality obtained by the reference SPIHT coder for a given targeted bit rate. In addition, the paper presents three-evaluation approaches (Objective, Subjective and Quantitative) based all on a quality score PS given by the objective visual quality metric named MWVDP which is an optimized version of Daly Visible Difference Predictor VDP. It incorporates the human visual system HVS properties, correlates largely well with the mean opinion score known as MOS and provides an important feature in image coding quality assessment. The visual coders EVIC/MEVIC are fully dependent to the HVS properties, from which, they include various masking effects and visual models. Based on this model, the visual coders weight the original wavelet coefficients and reshape their spectrum to optimize the perceptual quality coding for a given observation distance and bit rate. The visual weighting model processes within all wavelet sub-bands: 1) the contrast sensitivity filter CSF to mask invisible frequencies, 2) the threshold elevation to correct the luminance and elevate the contrast, 3) the Just Detectable Distortions JND to quantize visually the wavelet coefficients according to their corresponding thresholds. The visual coder EVIC and its optimized version MEVIC have the same software complexity as their reference SPIHT. However, they perform qualitatively and quantitatively excellent results experimentation and features improvement either in image coding and quality assessing.

Improving the transfer learning performances in the classification of the automotive traffic roads signs

“…Indeed, training the network on so few images is dangerous since the risk of overfitting, it's very important point in Deep Learning. Also, if the new images look like the old ones, then they can be represented by the same features [9].…”