In Artificial Intelligence, it is important to have some measure of the quality of the data we are trying to use. For instance, if we want to classify a cervix image according to a degree of cancer, how do we know if that image follows the acquisition protocol and can be used for diagnosing the patient [1] so that we can trust our prediction? In another domain, if we want to classify the topic of a given text (which may be written in colloquial language), how do you know if the text is useful/has good quality?

Image quality estimation is a tough problem since quality is dependent on the user and the task itself. This is of special interest when we measure quality according to several factors and use it to make decisions (e.g. efficiency, durability, price of something).

We have previously faced the issue of objective quality estimation in several areas (image/text) and industries (TelCo, healthcare). In this blog post, we will discuss some alternatives from the Data Science community, as well as a methodology we have successfully used in the past to estimate data quality without supervision on the quality task.

What’s out there?

Quality Assessment research’s goal, up until now, has been to make quality predictions that are in agreement with the opinion of human observers. As such, quality measurement is typically done using labels, which are very subjective – something of good quality to an observer can be of bad quality to another.

For instance, in [2], Convolutional Neural Networks are used for predicting subjective image quality in databases such as LIVE Image Quality Assessment Database, using human-assessed quality labels.

There are also some unsupervised approaches [3] that group image quality metrics into several groups, such as compression artifacts, image noise, color artifacts, blur, and distortions. However, these approaches focus on a “technical” view of quality (signal quality) instead of a semantic/human perception. In the text domain [4], a set of automatic rules is defined for Software Documentation Quality which is mostly based on human readability and ease of implementation.

All these previously mentioned approaches have been applied with some success, but most of them are focused on quantifying data acquisition errors. In some domains, such as the medical domain, it is important to have semantic knowledge of quality. As an example, in retinal fundus images, a common metric is the Image Structure Clustering (ISC). ISC assesses a correct distribution of pixel intensities corresponding to the relevant anatomical structures present in the retina [5].

However, creating such metrics and having them accepted by the researchers’ community, for each area, with lots of domain-specific rules, is not feasible. Besides, most methods for quality measurement are focused on a specific modality: images.

In this blog post, we propose a way to objectively measure the quality of any modality (image, text, tabular data), for both classification and regression problems, and give a few use cases in which we can use this.

General Methodology

Let’s suppose we are trying to classify a given sample (e.g. an image or a document), and we want to objectively measure the quality of that sample. 

We can consider that a good quality sample is the one reaches high scores in a secondary target task (e.g., classifying samples’ class, properly estimating a regression value, segmenting the image, etc.).

By modeling quality as task-specific, we have some semantic knowledge of quality for a given domain, but that is generalizable for any problem.

For instance, in Customer Service, we can measure how informative a client’s technical issue description is for understanding what happened with the client, or for understanding how close he is to churn (leaving the service). For medical images, we can measure how good an image is for classifying a certain form of cancer (and this can depend on illumination conditions, domain-specific artifacts, etc.). 

As such, the first step is having a model that can estimate the probability of that sample belonging to a certain class (discriminative model), or a regression model that estimates a continuous output.

We can then calculate the error associated with that estimation with the metric of interest. For instance, as  $ 1 - Probability(True class) $ in a classification task or by the absolute error/squared error in a regression task. By saving the out-of-fold predictions, we can have an error associated with each instance. 

To automatically estimate sample quality, we then need a second model that tries to predict the error using the sample as an input. 

Samples with higher error as an output of the second model will have lower quality, and therefore be more prone to having worse performance as an output of the first model. 

Image Quality in Fashion-MNIST 

Fashion-MNIST is a dataset of 60k 28×28 grayscale images of 10 fashion categories, along with a test set of 10,000 images, available in Keras’ datasets. We chose two classes to illustrate our approach for quality estimation (T-shirt/top vs trousers). How can we estimate the quality of this binary classification problem? 

Initially, both train and test splits were concatenated, as we intend to calculate out of fold errors. 

For the sake of simplicity, we introduced random zoom-in with a scale factor that can be manipulated to simulate the aggressiveness of our approach (in this case, a range of 0.5-0.9 was used from Keras’ ImageDataGenerator). The higher the scale factor, the more disrupted the image can be. With too low scale factors, the image might not have changes at all. On the other hand, if the scale factor is too high, the images lose their meaning. However, in a real-life setting you don’t need to apply data augmentation. Instead, you can simply use your original images with quality variations as input.

Below, 8 images were randomly chosen before and after adding noise. 

Model 1: Classifying samples 

We used a standard VGG-like CNN for classifying the samples.

The out-of-fold sample error was calculated using 5-Fold Cross-Validation. 

The next figure shows the Kernel Density Estimation of the model’s error. It can be seen that the model managed to obtain a low error on most of the samples, since this is a binary classification problem which uses a relatively simple dataset, but there are some samples with high error. 

Model 2: Estimating sample error

From the out-of-fold errors, we attempted to create a model which could calculate sample error. The out-of-fold samples were randomly split into 30% test and 70% train and a similar architecture was applied, also with a sigmoid activation and binary cross-entropy loss. MSE was not used as binary cross-entropy was found to have better convergence, and our output is between 0 and 1. 

The Spearman Rank was 44%, showing that we can properly estimate the classification error and its monotonicity having only an image. 

Below, the images the model considered to have the worst quality (highest estimated error) are shown. 

It can be seen that these are images with bad acquisition conditions, and with misleading shapes. Even the human eye has some difficulty in distinguishing both classes (top/trousers vs pants) in some of the images. 

Next, the images the model considered to have the best quality (lowest estimated error) are shown. Note that since this is an unbalanced regression problem, most of the images will have low error and the example shown below is not fully representative of the whole “good quality” examples. 

The higher quality images  – t-shirts – are the ones who follow a standard “image acquisition protocol” and follow overall image patterns.

We then performed another experiment, by removing the images with artificial data augmentation and re-training the whole process. In this new process, the images with highest error are the following:

Once again, lower quality images had a similar behavior to the previous example. On the other hand, the images with lowest error are the following:

These images are also very similar to the ones presented below. The model managed to ignore the patterns of the t-shirts’ drawings, and assumed the quality was good if the visible shape followed the “standards” of low-error images in the training set.

Discussion

We have managed to create a domain-independent approach for estimating sample error in classification problems without relying on subjective human annotations. 

In the literature, three groups of approaches were previously shown: supervised, unsupervised and domain-specific. We can frame our approach in a weakly-supervised approach, since we have a proxy of quality labels which is the task itself. 

While the proposed approach has the main advantage of not requiring human labels, it has a few disadvantages. We can have a low error in some cases, which can mislead the model into thinking that a bad image has high quality, or vice-versa. Some cases in which this can occur are:

  • Low-quality class labels
  • Areas with low density (i.e. less frequent types of images) which the model wasn’t able to learn correctly 
  • Cases which are difficult to learn and could have issues in the future (such as visually degraded images), but the model was good enough to discriminate them from the other class

There are alternative approaches for measuring quality which are not described here. 

Since this is not a generative model, we cannot fix the problem of possibly having high error in low-density areas with good quality data, since we do not know if we are in such areas. We can either pair our current model with another generative model or think about a different architecture.  

For using generative models, we can frame quality assessment as an anomaly detection problem. For instance, assuming the dataset has few bad-quality samples, we can learn its distribution. The lower the density of new samples, the worse its quality will be (or the more non-compliant it will be).  

References

[1]  Fernandes, Kelwin, Jaime S. Cardoso, and Jessica Fernandes. “Transfer learning with partial observability applied to cervical cancer screening.” Iberian conference on pattern recognition and image analysis. Springer, Cham, 2017.

[2] Po, L. M., Liu, M., Yuen, W. Y., Li, Y., Xu, X., Zhou, C., … & Luk, H. T. (2019). A Novel Patch Variance Biased Convolutional Neural Network for No-Reference Image Quality Assessment. IEEE Transactions on Circuits and Systems for Video Technology, 29(4), 1223-1229.

[3] D. Temel and G. AlRegib, Perceptual image quality assessment through spectral analysis of error representations, Signal Processing: Image Com-munication, Volume 70, 2019, Pages 37-46,ISSN 0923-5965

[4] Chen, M., & He, Y. (2017). Exploration on Automated Software Requirement Document Readability Approaches.

[5] Costa, P., Galdran, A., Meyer, M. I., Abràmoff, M. D., Niemeijer, M., Mendonça, A. M., & Campilho, A. (2017). Towards adversarial retinal image synthesis. arXiv preprint arXiv:1701.08974.