Distortion-based filtering for image classification
Methods and systems for image filtering include detecting a distortion level of input images, using a distortion detection model that is trained using confidence values generated by a pre-trained image classifier with a set of distorted training images. An analysis is performed on input images having a detected distortion level that is lower than a threshold, with input images having an above-threshold detected distortion level being filtered out.
1. A computer-implemented method for image filtering, comprising:
detecting a distortion level of input images, using a distortion detection model that is trained using confidence values generated by a pre-trained image classifier with a set of distorted training images;
performing analysis on input images having a detected distortion level that is lower than a threshold, with input images having an above-threshold detected distortion level being filtered out; and
adjusting an image capture setting responsive to a determination that the input image has an above-threshold distortion level.
2. The computer-implemented method of claim 1 , further comprising training the distortion detection model using the set of distorted training images, each distorted training image having an associated target score that is based on the confidence values generated by the pre-trained image classifier.
3. The computer-implemented method of claim 2 , further comprising generating the set of distorted training images by performing a plurality of distortion operations on original training images.
4. The computer-implemented method of claim 3 , wherein generating the set of distorted training images further includes generating a target score for each respective distorted training image by aggregating confidence scores from a plurality of pre-trained image classifiers using the respective distorted training image as input.
5. The computer-implemented method of claim 1 , further comprising training the distortion detection model using a set of distorted training images, each distorted training image having an associated target score that is based on a distance between a classifier softmax output for the distorted training image and a classifier softmax output for an original undistorted version of the distorted training image.
6. The computer-implemented method of claim 1 , wherein the image capture setting is selected from the group consisting of a camera frame rate and a brightness.
7. The computer-implemented method of claim 1 , wherein the distortion detection model is implemented as a neural network model that includes a feature extractor with a plurality of branches of convolutional layers and a regressor with fully connected layers.
8. The computer-implemented method of claim 7 , wherein the plurality of branches include convolutional layers of different kernel sizes.
9. The computer-implemented method of claim 8 , wherein the feature extractor concatenates outputs of the plurality of branches to generate a feature extractor output.
10. A computer-implemented method for image filtering, comprising:
performing a plurality of distortion operations on a set of original training images to generate distorted training images corresponding to the original training images;
training a neural network of a distortion detection model using the set of original training images and the distorted training images corresponding to the original training images, based on confidence values generated by a pre-trained image classifier;
detecting a distortion level of input images, using the distortion detection model;
performing analysis on input images having a detected distortion level that is lower than a threshold, with input images having an above-threshold detected distortion level being filtered out; and
adjusting an image capture setting responsive to a determination that the input image has an above-threshold distortion level.
11. A system for image filtering, comprising:
a hardware processor; and
a memory that stores a computer program, which, when executed by the hardware processor, causes the hardware processor to:
detect a distortion level of input images, using a distortion detection model that is trained using confidence values generated by a pre-trained image classifier with a set of distorted training images;
perform analysis on input images having a detected distortion level that is lower than a threshold, with input images having an above-threshold detected distortion level being filtered out; and
adjust an image capture setting responsive to a determination that the input image has an above-threshold distortion level.
12. The system of claim 11 , wherein the image capture setting is selected from the group consisting of a camera frame rate and a brightness.
13. The system of claim 11 , wherein the computer program further causes the hardware processor to train distortion detection model using the set of distorted training images, each distorted training image having an associated target score that is based on the confidence values generated by the pre-trained image classifier.
14. The system of claim 13 , wherein the computer program further causes the hardware processor to generate the set of distorted training images by performing a plurality of distortion operations on original training images.
15. The system of claim 14 , wherein the computer program further causes the hardware processor to generate a target score for each respective distorted training image by aggregating confidence scores from a plurality of pre-trained image classifiers using the respective distorted training image as input.
16. The system of claim 11 , wherein the computer program further causes the hardware processor to train the distortion detection model using a set of distorted training images, each distorted training image having an associated target score that is based on a distance between a classifier softmax output for the distorted training image and a classifier softmax output for an original undistorted version of the distorted training image.
17. The system of claim 11 , wherein the distortion detection model is implemented as a neural network model that includes a feature extractor with a plurality of branches of convolutional layers and a regressor with fully connected layers.
18. The system of claim 17 , wherein the plurality of branches include convolutional layers of different kernel sizes.