Accelerating OpenCV Image Matching with CUDA: CPU vs GPU Performance

Background

In automated testing, using image recognition to locate UI controls is a common requirement, and the response speed of the HTTP API that provides image recognition is critical.

Problem

This article focuses on accelerating the relevant OpenCV APIs; other server‑side optimization techniques are out of scope. By default, opencv‑python does not include GPU (CUDA) support, so you must compile it yourself. The article uses OpenCV 3.2.0 with CUDA 8.0, as recommended by the official documentation; other versions may cause compilation failures.

Example – Image Matching

# coding=utf-8
import cv2
import time

def match_test():
    target = cv2.imread("./target.png")
    template = cv2.imread("./template.jpg")
    result = cv2.matchTemplate(target, template, cv2.TM_CCOEFF_NORMED)
    minVal, maxVal, minLoc, maxLoc = cv2.minMaxLoc(result)
    h, w = template.shape[:-1]
    if maxVal > 0.5:
        middle_point = (int(maxLoc[0] + w / 2), int(maxLoc[1] + h / 2))
        return middle_point
    else:
        return None

if __name__ == '__main__':
    num = 100
    begin = time.time()
    for i in range(num):
        match_test()
    print((time.time() - begin) / num)

Result

CPU: 0.299 seconds per call

GPU: 0.181 seconds per call

Improvement: 39.4% faster using GPU acceleration.