Table 1
Learning–Testing Strategy for a Preliminary Experiment
Based on Blur-Level Classification Including a Single Representative
Scene for Each of Five Types in Both the Learning and the Testing Sets
| Image Type
| Learn
| Test
|
|---|
| (a) Animal | A | B |
| (b) Human face | A | B |
| (c) Natural landscape | A | B |
| (d) Architecture | A | B |
| (e) Other
| A
| B
|
Table 2
Classification Results Detailing Error Occurrences for
Each Degradation Level
| Image Quality
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (o) Very good | 5 | | | | |
| (1) Good | | 5 | | | |
| (2) Fair | | | 4 | (d) | |
| (3) Poor | | | | 4 | (a) |
| (4) Very poor
| | | | (e)
| 4
|
Table 3
Learning–Testing Strategy for a Preliminary Experiment
Based on Blur-Level Recognition Including a Single Representative Scene
for Only Three of the Five Types in the Learning Set and All Five Types
in the Testing Set
| Image Type
| Learn
| Test
|
|---|
| (a) Animal | A | B |
| (b) Human face | | A, B |
| (c) Natural landscape | | A, B |
| (d) Architecture | A | B |
| (e) Other
| A
| B
|
Table 4
Recognition Results
| Image Quality
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (o) Very good | 7 | | | | |
| (1) Good | | 7 | | | |
| (2) Fair | | | 6 | (c) | |
| (3) Poor | | | (e) | 5 | (a) |
| (4) Very poor
| | | | (e)
| 6
|
Table 5
Error Rate Fraction for Blur-Level Classification Showing
the Number of Cycles in the Training
| Error Rate Fractions
| Number of Training Sets
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| Ring-only cycles | 10,000 | 20,000 | 30,000 | 40,000 | 50,000 |
| Image | 25/125 | 15/100 | 8/75 | 4/50 | 1/25 |
| Edge profile | 52/125 | 27/100 | 13/75 | 9/50 | 5/25 |
| Image and edge profile | 23/125 | 14/100 | 6/75 | 3/50 | 0/25 |
| Wedge-only cycles | 100,000 | 100,000 | 350,000 | 400,000 | 500,000 |
| Image | 100/125 | 70/100 | 50/75 | 20/50 | 10/25 |
| Edge profile | 99/125 | 72/100 | 55/75 | 23/50 | 11/25 |
| Image and edge profile | 98/125 | 70/100 | 52/75 | 22/50 | 11/25 |
| Ring–wedge cycles | 10,000 | 20,000 | 30,000 | 40,000 | 50,000 |
| Image | 30/125 | 16/100 | 10/75 | 3/50 | 1/25 |
| Edge profile | 50/125 | 23/100 | 10/75 | 6/50 | 3/25 |
| Image and edge profile | 26/125
| 15/100
| 8/75
| 4/50
| 1/25
|
Table 6
Blur-Level Classification Accuracy for Only Ring Data from
Both Gray-Scale Imagery and Corresponding Edge Profiles for a Data Set
of 250 Separate Images in the Testing Set with Ten Errors
| Image Quality
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (o) Very good | 49 | 1 | | | |
| (1) Good | | 49 | 1 | | |
| (2) Fair | | 2 | 48 | 1 | |
| (3) Poor | | | 2 | 46 | 2 |
| (4) Very poor
| | | | 2
| 48
|
Table 7
Blur-Level Classification Accuracy for Only Wedge Data
from Both Gray-Scale Imagery and Corresponding Edge Profiles for a Data
Set of 250 Separate Images in the Testing Set with 59 Errors
| Image Quality
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (o) Very good | 49 | 1 | | | |
| (1) Good | | 47 | 3 | | |
| (2) Fair | | 5 | 36 | 9 | |
| (3) Poor | | | 11 | 15 | 24 |
| (4) Very poor
| | | 4
| 2
| 44
|
Table 8
Blur-Level Classification Accuracy for Both Ring and Wedge
Data from Gray-Scale Imagery and Corresponding Edge Profiles for a Data
Set of 250 Separate Images in the Testing Set with 18 Errors
| Image Quality
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (o) Very good | 49 | 1 | | | |
| (1) Good | | 49 | 1 | | |
| (2) Fair | | 1 | 47 | 2 | |
| (3) Poor | | | 3 | 42 | 5 |
| (4) Very poor
| | | 3
| 2
| 45
|
Table 9
Blur-Level Classification Accuracy Using Four Common Image
Objective Image Quality Measures from Gray-Scale Imagery for a Data Set
of 250 Separate Images in the Testing Set with an Overall Accuracy of
53%
| Image Quality
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (o) Very good | 50 | | | | |
| (1) Good | | 35 | 9 | 2 | 4 |
| (2) Fair | | 17 | 15 | 4 | 14 |
| (3) Poor | | 13 | 10 | 7 | 20 |
| (4) Very poor
| | 10
| 8
| 5
| 27
|
Table 10
JPEG Quality Classification Accuracy for Only Ring Data
from Both Gray-Scale Imagery and Corresponding Edge Profiles for a Data
Set of 250 Separate Images in the Testing Set with 20 Errors
| Image Quality (q)
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (100) | Very good | 49 | 1 | | | |
| (40) | Good | 1 | 45 | 3 | 1 | |
| (10) | Fair | 1 | 3 | 44 | 2 | |
| (5) | Poor | | | 3 | 44 | 3 |
| (1) | Very poor
| | | | 2
| 48
|
Table 11
JPEG Quality Classification Accuracy for Only Wedge Data
from Both Gray-Scale Imagery and Corresponding Edge Profiles for a Data
Set of 250 Separate Images in the Testing Set with 66 Errors
| Image Quality (q)
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (100) | Very good | 45 | 4 | 1 | | |
| (40) | Good | 9 | 37 | 2 | 2 | |
| (10) | Fair | 2 | 4 | 34 | 10 | |
| (5) | Poor | | 5 | 6 | 28 | 11 |
| (1) | Very poor
| | 1
| 2
| 7
| 40
|
Table 12
JPEG Quality Classification Accuracy for Both Ring and
Wedge Data from Gray-Scale Imagery and Corresponding Edge Profiles for
a Data Set of 250 Separate Images in the Testing Set with 13 Errors
| Image Quality (q)
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (100) | Very good | 50 | | | | |
| (40) | Good | | 50 | | | |
| (10) | Fair | | 3 | 45 | 2 | |
| (5) | Poor | | | 2 | 45 | 3 |
| (1) | Very poor
| | | | 3 | 47 |
Table 13
JPEG Fidelity Classification Accuracy for Both Ring and
Wedge Data from Error Images between Degraded Images and Corresponding
High-Quality Originals for a Data Set of 250 Separate Images in the
Testing Set with Five Errors
| Image Quality (q)
| Maximum Valued Output Neuron
|
|---|
| 1
| 2
| 3
| 4
| 5
|
|---|
| (100) | Very good | 50 | | | | |
| (40) | Good | | 50 | | | |
| (10) | Fair | | 2 | 47 | 1 | |
| (5) | Poor | | | 1 | 48 | 1 |
| (1) | Very poor
| | | | | 50
|
Table 14
Summary of Results for Automatic Image Quality Assessment
Using the All-Digital Ring–Wedge Detector
| Section/Subsection
| Description
| Results
|
|---|
| 3.A | Blur-level sorting
Preliminary study | Good sorting results independent of scene content. Best to consider each image type in training. |
| 3.B | Blur-level sorting
System refinement | Ring data most important in making assessments. Improved sorting results by inclusion of data from edge profiles of individual images. Best to consider several image examples per image type in training. |
| 3.C | Blur-level sorting
System evaluation | Demonstrates excellent classification of a common linear degradation type: |
| | | 96% accuracy (ring data only) |
| | | 76% accuracy (wedge data only) |
| | | 92% accuracy (ring and wedge data) |
| 4.A | JPEG study
Quality sorting | Demonstrates excellent classification of an important nonlinear degradation type. Wedge data useful for characterizing the severity of artifacts in the scene: |
| | | 92% accuracy (ring data only) |
| | | 73% accuracy (wedge data only) |
| | | 95% accuracy (ring and wedge data) |
| 4.B | JPEG study
Fidelity sorting | Demonstrates a novel technique for evaluating image fidelity: |
| | | 99% accuracy (ring and wedge data) |
| 5 | Localized ring–wedge transform
| Effectively combines both image-domain and spatial-transform-domain information. Effectively estimates local fidelity with data from the degraded scene alone, without explicit information about the original image.
|
Add to BibSonomy