Scalable parallel extraction of linear features on MP-2

Scalable parallel algorithms to extract linear features from a grey level image are shown. The MasPar MP-2 is modeled as a p × p processor array. Given an n × n image, the time complexity of the algorithm which performs linear feature extraction is O(km²n ²/p² + n), where the size of the convolution masks is m × m and k denotes the number of masks used. The overall system is processor-time optimal and is scalable in the range 1 ⩽ p ² ⩽ P, where P = Θ(km²n). Based on this, a low-level vision system to extract linear features is implemented on MP-2, and the performance results are reported. Given a 512 × 512 grey level image as input, the edge detection task consisting of applying six 5 × 5 convolution masks takes less than 118 msec on a 4 K processor MP-2. This is a 540 fold speedup compared with the serial implementation on a SUN SPARC-400 workstation which takes 64 sec. Extraction of linear features using symbolic approaches, which includes edge detection, thinning, linking, contour tracing, and linear approximation, takes less than 1.5 sec. These implementations are scalable with respect to the machine size. Results on various sizes of images using various sizes of MP-2 are also reported. The bottlenecks and suitability of SIMD machines for low level vision processing are identified