Image quality assessment using adaptive non-overlapping mean estimation
Methods, apparatus, systems and articles of manufacture (e.g., physical storage media) to assess image quality using adaptive non-overlapping mean estimation are disclosed. Example image quality assessment methods disclosed herein include replacing respective blocks of pixels of a first image with mean values of the respective blocks of pixels to determine a second image having a smaller size than the first image. Disclosed example image quality assessment methods also include determining a vector of features for the second image. Disclosed example image quality assessment methods further include applying the vector of features to a neural network, and classifying a quality of the first image based on an output of the neural network.
1. An apparatus to perform image quality assessment, the apparatus comprising:
a training module to train a neural network, based on a set of training images, to classify quality of a first image;
an image determiner to replace respective blocks of pixels of the first image with mean values of the respective blocks of pixels to determine a second image having a smaller size than the first image;
a feature extractor to determine a vector of features including a blur feature value for the second image and to apply the vector of features to the neural network after the neural network has been trained, the feature extractor to determine the blur feature value based on an amount the second image differs from a version of the second image filtered with a blurring filter in horizontal and vertical directions of the second image; and
an image classifier to classify a quality of the first image based on an output of the neural network, the training module, the image determiner, the feature extractor, and the image classifier implemented by hardware circuitry or at least one processor.
2. The apparatus of claim 1 , wherein the respective blocks of pixels of the first image form adjacent blocks of pixels that are non-overlapping.
3. The apparatus of claim 1 , wherein to determine the second image, the image determiner is further to:
segment the first image into the respective blocks of pixels, the respective blocks of pixels having a first number of pixels in a horizontal direction and a second number of pixels in a vertical direction; and
for a first one of the respective blocks of pixels:
average values of the pixels in the first one of the respective blocks of pixels to determine a first mean value corresponding to the first one of the respective blocks of pixels; and
set a first pixel of the second image to be the first mean value, the first pixel at a first position in the second image that is to represent the first one of the respective blocks of pixels of the first image.
4. The apparatus of claim 3 , wherein the first number of pixels is two pixels, the second number of pixels is two pixels, the respective blocks of pixels of the first image each include four pixels, and the second image is one-fourth the size of the first image.
5. The apparatus of claim 1 , wherein to classify the quality of the first image, the image classifier is to perform a single classification of the quality of the first image by:
classifying the first image as acceptable for an image auditing application when the output of the neural network satisfies a threshold; and
classifying the first image as unacceptable for the image auditing application when the output of the neural network does not satisfy the threshold.
6. The apparatus of claim 1 , wherein to classify the quality of the first image, the image classifier is to compare a plurality of outputs of the neural network to a plurality of thresholds to perform multiple classifications of the quality of the first image.
7. An image quality assessment method comprising:
training, by executing an instruction with a processor, a neural network, based on a set of training images, to classify quality of a first image;
replacing, by executing an instruction with the processor, respective blocks of pixels of the first image with mean values of the respective blocks of pixels to determine a second image having a smaller size than the first image;
determining, by executing an instruction with the processor, a vector of features including a blur feature value for the second image, the blur feature value determined based on an amount the second image differs from a version of the second image filtered with a blurring filter in horizontal and vertical directions of the second image;
applying the vector of features to the neural network after the neural network has been trained; and
classifying, by executing an instruction with the processor, a quality of the first image based on an output of the neural network.
8. The method of claim 7 , wherein the respective blocks of pixels of the first image form adjacent blocks of pixels that are non-overlapping.
9. The method of claim 7 , wherein the replacing of the respective blocks of pixels of the first image with mean values of the respective blocks of pixels to determine the second image includes:
segmenting the first image into the respective blocks of pixels, the respective blocks of pixels having a first number of pixels in a horizontal direction and a second number of pixels in a vertical direction; and
for a first one of the respective blocks of pixels:
averaging values of the pixels in the first one of the respective blocks of pixels to determine a first mean value corresponding to the first one of the respective blocks of pixels; and
setting a first pixel of the second image to be the first mean value, the first pixel at a first position in the second image that is to represent the first one of the respective blocks of pixels of the first image.
10. The method of claim 9 , wherein the first number of pixels is two pixels, the second number of pixels is two pixels, the respective blocks of pixels of the first image each include four pixels, and the second image is one-fourth the size of the first image.
11. The method of claim 7 , further including converting the second image to a grayscale image before determining the vector of features for the second image.
12. The method of claim 7 , wherein the classifying of the quality of the first image includes performing a single classification of the quality of the first image by:
classifying the first image as acceptable for an image auditing application when the output of the neural network satisfies a threshold; and
classifying the first image as unacceptable for the image auditing application when the output of the neural network does not satisfy the threshold.
13. The method of claim 7 , wherein the classifying of the quality of the first image includes comparing a plurality of outputs of the neural network to a plurality of thresholds to perform multiple classifications of the quality of the first image.
14. A tangible computer readable storage medium comprising computer readable instructions that, when executed, cause a processor to at least:
train a neural network, based on a set of training images, to classify quality of a first image;
replace respective blocks of pixels of the first image with mean values of the respective blocks of pixels to determine a second image having a smaller size than the first image;
determine a vector of features including a blur feature value for the second image, the blur feature value determined based on an amount the second image differs from a version of the second image filtered with a blurring filter in horizontal and vertical directions of the second image;
apply the vector of features to the neural network after the neural network has been trained; and
classify a quality of the first image based on an output of the neural network.
15. The tangible computer readable storage medium of claim 14 , wherein the respective blocks of pixels of the first image form adjacent blocks of pixels that are non-overlapping.
16. The tangible computer readable storage medium of claim 14 , wherein to determine the second image, the instructions, when executed, further cause the processor to:
segment the first image into the respective blocks of pixels, the respective blocks of pixels having a first number of pixels in a horizontal direction and a second number of pixels in a vertical direction; and
for a first one of the respective blocks of pixels:
average values of the pixels in the first one of the respective blocks of pixels to determine a first mean value corresponding to the first one of the respective blocks of pixels; and
set a first pixel of the second image to be the first mean value, the first pixel at a first position in the second image that is to represent the first one of the respective blocks of pixels of the first image.
17. The tangible computer readable storage medium of claim 16 , wherein the first number of pixels is two pixels, the second number of pixels is two pixels, the respective blocks of pixels of the first image each include four pixels, and the second image is one-fourth the size of the first image.
18. The tangible computer readable storage medium of claim 14 , wherein the instructions, when executed, further cause the processor to convert the second image to a grayscale image before determining the vector of features for the second image.
19. The tangible computer readable storage medium of claim 14 , wherein to classify the quality of the first image, the instructions, when executed, cause the processor to perform a single classification of the quality of the first image by:
classifying the first image as acceptable for an image auditing application when the output of the neural network satisfies a threshold; and
classifying the first image as unacceptable for the image auditing application when the output of the neural network does not satisfy the threshold.
20. The tangible computer readable storage medium of claim 14 , wherein to classify the quality of the first image, the instructions, when executed, cause the processor to compare a plurality of outputs of the neural network to a plurality of thresholds to perform multiple classifications of the quality of the first image.