Retinal Image Registration Using Partial Intensity Invariant Feature Descriptor and Redundant Keypoints Elimination Techniques

Retinal image registration is crucial for improving the accuracy of diagnosing and monitoring retinal diseases. This paper proposes a specific feature region technique for retinal image registration, involving five steps for accurate registration. The technique combines area and feature-based methods, including extracting the retinal vascular tree, detecting distinctive points, eliminating redundant keypoints, and matching features for image registration using affine transformation modes. First, the retina’s vascular tree is extracted using a Top-Hat operation and optimal thresholding technique. Next, the Harris-PIIFD detector identifies key points in the binary image, and redundant keypoints are removed to reduce computational load. Finally, bilateral matching and best-bin-first algorithms compute the similarity matrix for registering the images. If the image pair is accepted, the points are controlled using the simplest affine transformation modes for the highest registration success rate. The simulation results on 134 pairs of FIRE datasets demonstrate the effectiveness and robustness of the proposed algorithm. The experimental results for the proposed method are satisfactory. The result obtained for the proposed approach for retinal image registration is precision 0.98240, recall 0.98312, RMSE 0.01280, ERR 0.01716, and matching score 0.98284, with the computational time taken 3.01s. This hybrid image registration approach is an efficient and reliable tool for retinal image registration, leading to more accurate diagnosis and monitoring of retinal diseases.