Thesis-NoiseDectection-results/NoiseDectection-attributes/RESULTS.md

NoiseDetection-metrics

Description

All results obtained from NoiseDetection-metrics project in order to classify noisy or not noisy synthesis images obtained from monte carlo rendering process.

Results

1.SV_with_SVM

Description

Study of all SV values obtained from SVD compression of images and trying to classify them (noisy/not noisy image).

Parameters

• Interval chosen from SV (begin and end)
• Number of components from SV
• Kind of normalization (3 differents way)
• Number of zones taken ramdomly in training set from the 16 available of an image
• The features used from pool of features computed from SV

Pool of features

• lab (Use of luminance canal obtained from Lab)
• mscn (Mean Substracted Coeffiecients Normalized)
• low_bits_2 (only 2 bits taken from from lab)
• low_bits_3 (only 3 bits taken from from lab)
• low_bits_4 (only 4 bits taken from from lab)
• low_bits_5 (only 5 bits taken from from lab)
• low_bits_6 (only 6 bits taken from from lab)
• low_bits_4_shifted_2 (only 4 bits shifted by 2 taken from from lab)

2.SV_Statistics

Description

Study of statistics obtained from SV vector obtained from SVD compression.

Computed statistics are:

• Mean
• Median
• Standard deviation
• Variance
• Percentile
• Area under SV curve

Parameters

• Kind of normalization (3 differents way)
• Number of zones taken ramdomly in training set from the 16 available of an image
• The features (here computed statistics) used from pool of features computed from SV

Pool of features

• sub_blocks_stats (all stats)
• sub_blocks_area (only area under curves from sub blocks)
• sub_blocks_stats_reduced
• sub_blocks_area_normed

3.SV_Correlation

Description

Study of correlation (using pearson correlation coefficient) between each components of SV vector on the whole dataset (all scences data).

Parameters

• $n$, the number of highest or lowest correlated components to keep
• Kind of normalization (3 differents way)
• Number of zones taken ramdomly in training set from the 16 available of an image

Pool of features

• lab (Use of luminance canal obtained from Lab)
• mscn (Mean Substracted Coeffiecients Normalized)
• low_bits_2 (only 2 bits taken from from lab)
• low_bits_3 (only 3 bits taken from from lab)
• low_bits_4 (only 4 bits taken from from lab)
• low_bits_5 (only 5 bits taken from from lab)
• low_bits_6 (only 6 bits taken from from lab)
• low_bits_4_shifted_2 (only 4 bits shifted by 2 taken from from lab)

4.MSCN_stats

Description

Study of statistics on MSCN matrix obtained on synthesis image. Only variance is computed on sub blocks.

Parameters

• Kind of normalization (3 differents way)
• Number of zones taken ramdomly in training set from the 16 available of an image
  

Pool of features

Note that mscnvar$n$ is the number of sub blocks.

• mscn_var_4
• mscn_var_16
• mscn_var_64
• mscn_var_16_max (data is sorted from max variance to min)
• mscn_var_64_max (data is sorted from max variance to min)
  

5.Data_separation

Description

Use of different way to train data:

• using all data
• using splited data from human threshold
• using only data close to human threshold

Parameters

• Kind of normalization (3 differents way)
• Number of zones taken ramdomly in training set from the 16 available of an image

Pool of features

• lab (Use of luminance canal obtained from Lab)
• mscn (Mean Substracted Coeffiecients Normalized)
• low_bits_2 (only 2 bits taken from from lab)
• low_bits_3 (only 3 bits taken from from lab)
• low_bits_4 (only 4 bits taken from from lab)
• low_bits_5 (only 5 bits taken from from lab)
• low_bits_6 (only 6 bits taken from from lab)
• low_bits_4_shifted_2 (only 4 bits shifted by 2 taken from from lab)
• ica_diff
• svd_trunc_diff
• ipca_diff
• svd_reconstruct
• highest_sv_std_filters (use of different filters on images and SVD computed on it)
• lowest_sv_std_filters (use of different filters on images and SVD computed on it)

NoiseDetection-26-attributes

Description

All results obtained from NoiseDetection-26-attributes project in order to classify noisy or not noisy synthesis images obtained from monte carlo rendering process.

Results

1.FirstResults

Description

Study of statistics features extracted from filters applied on synthesis images in order to detect noise.

Parameters

• Kind of normalization (3 differents way)
• Number of zones taken ramdomly in training set from the 16 available of an image

Pool of features

• filters_statistics (statistics obtained from applied fitlers)