Illumination-Independent Object-Based Change Detection in Linear Complexity
This research presents an object-based change detection algorithm for pairs of registered gray-level images captured under substantially different imaging conditions. The central innovation is that the decision is not based on direct gray-level subtraction, which is highly sensitive to lighting variation, but on significant blob extraction, boundary saliency, and gradient-distribution matching.
First Stage — Connectivity Along Gray Levels
Each image is transformed into a sequence of thresholded binary images I(t), where connected components evolve as the threshold changes. Significant blobs are detected as components whose boundary-weight function reaches a meaningful local maximum. This makes the extraction process local and object-oriented rather than block-based.
Second Stage — Illumination-Agnostic Object Candidates
The algorithm extracts candidate blobs from both images and also from their negative images, allowing both dark and bright objects to be detected. Because the extraction relies on connectivity and gradient saliency along boundaries, the method remains robust even when one image is captured in daylight and the other under night or infrared conditions.
Third Stage — Boundary Saliency and Distribution
For each candidate object Oᵢ, the method compares the strength and distribution of gradient magnitudes along the same boundary coordinates in the two images. The saliency measure SM, saliency ratio SR, entropy-based distribution measure dst, and distribution ratio DSR jointly determine whether the object exists in only one image.
Fourth Stage — Object of Change Decision
The final change score combines boundary saliency and boundary-gradient distribution. A candidate is declared an object of change when its contour is salient in one image but does not exhibit a corresponding boundary structure in the other image. This formulation makes the algorithm largely agnostic to extreme illumination changes and suitable for real-time applications with almost linear complexity.