Rectification of Image Deformation using PSO-CNN approach

Abstract

Tasks like document scanning, object recognition, and automated monitoring are greatly hindered by geometric distortion in images, which is often due to faulty camera angles, lens aberrations, or scene perspective. This paper, through combination of classic image processing techniques with metaheuristic and soft computing methodologies, presents an adaptive hybrid solution for distortion correction.  Using Gaussian filtering and then Canny edge detection, images are preprocessed to highlight significant edges. Possible quadrilateral regions are identified by removing contours and using polygonal fitting to approximate them. These initial vertices are enhanced by the application of Particle Swarm Optimization (PSO). A fitness function based on geometric regularity, such aspect ratio and angle consistency, promises an optimum rectification mapping.

To reconstruct a geometrically rectified picture, the optimized corner points are then sent into the homography transformation.

With the purpose of increasing accuracy in scenarios like occlusion, shadowing, or poor illumination, a Convolutional Neural Network (CNN) is trained to predict the position of deformed picture vertices straight from raw pixel input. By incorporating learned spatial priors into the PSO's optimization, the CNN serves as a backup or fine-tuning technique for unpredictable detections. Experimental results signify that the proposed method outperforms traditional Hough or Radon-based correction algorithms in terms of pixel alignment, structural similarity, and correction precision, thus proving its robustness in a range of real-world circumstances. In applications including document analysis, remote sensing, and intelligent vision, this unified PSO-CNN system offers a scalable and training-efficient method for trustworthy geometric correction.