A survey of Identification and mitigation of Machine Learning algorithmic biases in Image Analysis

Bias in image-based AI systems is often invisible: Unlike applications with obvious demographic variables, biases in image analysis can be "indirectly represented in the input data, or can be unknown," making them harder to detect and address than biases in other AI applications.

Safety-critical industries lack adequate bias safeguards: Industrial and safety-critical applications using image analysis are "most of the times sailing blind" regarding potential biases, creating serious deployment risks that regulatory frameworks are now beginning to address.

Three-step approach is needed: The survey identifies that tackling algorithmic bias requires: (1) understanding how biases manifest in image systems, (2) developing methods to detect and expose hidden biases, and (3) implementing mitigation strategies—suggesting no single solution exists.

Regulatory pressure is mounting: New regulations "clearly address the issues opened by undesired biases in AI," indicating that organizations deploying image-based AI systems face both ethical and legal obligations to identify and mitigate these problems.

Real-world applications demonstrate feasibility: The paper includes "a practical remote sensing use-case of industrial Fairness," suggesting that bias identification and mitigation in image analysis is not merely theoretical but can be implemented in actual commercial applications.