A Low-Complexity Canny Edge Detector for High-Resolution Mobile Imaging

Edge detection is a foundational step in computer vision pipelines, enabling object recognition, scene understanding, and visual navigation. The classical Canny algorithm, while optimal in detection and localisation accuracy, imposes substantial computational overhead through two-dimensional Gaussian convolution, Euclidean gradient magnitude computation, global histogram-based threshold selection, and sequential non-maximum suppression and hysteresis passes, rendering it impractical for battery-operated mobile platforms without modification. This paper presents a low-complexity Canny edge detector optimised for high-resolution mobile imaging through four algorithmic substitutions: separable one-dimensional Gaussian smoothing reducing kernel operations from O(k²) to O(2k) per pixel;a three-direction gradient magnitude approximation eliminating square-root computation with error below 4%; a block-adaptive double-threshold scheme on 64×64 tiles replacing global histogram analysis with local mean estimation; and a unified non-maximum suppression and hysteresis pass merging two memory scans into one. Validated on the BSDS300 dataset and real-time webcam streams, the proposed detector achieves edge quality comparable to the OpenCV Canny baseline with approximately 33% reduction in processing time.