Core Concepts
Deep learning models trained for image age approximation may exploit content bias rather than solely relying on age-related features, posing challenges for explainability.
Abstract
This paper proposes a novel approach to evaluate the influence of image content on the performance of deep learning models trained for image age approximation. The key insights are:
Deep learning models, such as AlexNet and SRNet, trained on regular scene images tend to exploit image content rather than solely relying on age-related features like in-field sensor defects. This is demonstrated by evaluating the models' performance on different types of average images, where content is suppressed to varying degrees.
Preprocessing techniques like median filter residuals and constrained convolutional layers can help increase the signal-to-noise ratio (age signal to image content) and reduce the influence of content bias. However, image content still plays an important role in the inference of these models.
The proposed XAI method is validated using synthetic images, where content bias can be ruled out. The results show that the method can effectively distinguish whether a model is exploiting age-related features or image content.
The findings suggest that classical approaches for image age approximation, which directly exploit age-related traces like in-field sensor defects, may be more reliable than feature-learning based deep learning models at the current stage, as the latter are prone to content bias issues.
Stats
The average accuracy of the AlexNet model on the original input images is 0.94.
The average difference in accuracy between the original input images and the average images is 0.69 for the AlexNet model.
The average accuracy of the SRNet-cs model on the original input images is 0.95.
The average difference in accuracy between the original input images and the average images is 0.22 for the SRNet-cs model.
Quotes
"Deep neural networks can be considered as a 'black box'. For example, in the context of deep learning image age approximation, it is not evident that inference is based solely on detected age traces."
"Based on the results and in comparison to all variants examined, the image content is least involved in inference with the SRNet-cs."