Identifying Bias in Deep Neural Networks Using Image Transforms

TL;DR

This work tackles the problem of hidden biases in image classification benchmarks by treating CNNs as potential receivers of non-object signals. It introduces a transform-based framework using $ Fourier $, $ Wavelet $, and $ Median ext{-}filter $ operations applied to whole images to distinguish contextual information from background bias, without requiring background cropping. Across natural, synthetic, and mixed datasets, the transforms produce distinct bias signatures: Fourier generally reduces accuracy across the board, while wavelets and median filtering reveal or amplify background signals in biased datasets. The approach provides a practical tool for bias detection and model validation, with code available for reproducibility and potential bias correction via generative methods in future work.

Abstract

CNNs have become one of the most commonly used computational tool in the past two decades. One of the primary downsides of CNNs is that they work as a ``black box", where the user cannot necessarily know how the image data are analyzed, and therefore needs to rely on empirical evaluation to test the efficacy of a trained CNN. This can lead to hidden biases that affect the performance evaluation of neural networks, but are difficult to identify. Here we discuss examples of such hidden biases in common and widely used benchmark datasets, and propose techniques for identifying dataset biases that can affect the standard performance evaluation metrics. One effective approach to identify dataset bias is to perform image classification by using merely blank background parts of the original images. However, in some situations a blank background in the images is not available, making it more difficult to separate foreground or contextual information from the bias. To overcome this, we propose a method to identify dataset bias without the need to crop background information from the images. That method is based on applying several image transforms to the original images, including Fourier transform, wavelet transforms, median filter, and their combinations. These transforms were applied to recover background bias information that CNNs use to classify images. This transformations affect the contextual visual information in a different manner than it affects the systemic background bias. Therefore, the method can distinguish between contextual information and the bias, and alert on the presence of background bias even without the need to separate sub-images parts from the blank background of the original images. Code used in the experiments is publicly available.

Figures (17)

• Figure 1: X-rays images classified correctly by CNNs. a) Original X-ray, b) AlexNet, c) GoogleNet, d) VGG16, e) VGG19, f) ResNet18, g) ResNet50, h) ResNet101, i) Inception V3, j) InceptionResNet, k) DenseNet201, l) SqueezeNet, m) Xception, and n) CNN-X majeed2020issues
• Figure 2: A 20$\times$20 cropped background segment of a blank background part of the original image. That was done to all images in the dataset to create a new dataset of blank sub-images, as shown in Figure \ref{['fig_yale_cropped']}. When using just these seemingly blank parts of the images, the classification accuracy of numerous datasets was far higher than mere chance accuracy.
• Figure 3: Original images from Yale Faces B and the 20$\times$20 portion of the top left corner separated from each of the original images. The classification accuracy of the CNN was far higher than mere chance, showing the the CNN does not necessarily need to recognize the face in order to classify the images correctly.
• Figure 4: Original images from KVASIR and the 20$\times$20 portion of the top left corner separated from the original images sanchari2021.
• Figure 5: Classification accuracy of CNN models trained and tested on seemingly blank sub-images taken from image background of several common image benchmark datasets sanchari2021.