【项目实训】成绩识别录入系统--opencv实现图片矫正 （二）

---

一、图片校正实现

部分代码

if len(contours2) == 4:
    c4 = contours2.reshape(4, 2)
    if fclock[0]:
        print
        "c4=", c4
        c4list = c4.tolist()
        c4list.reverse()
        c4 = np.array(c4list)


    c5 = np.zeros((4, 2), dtype=np.float32)

    minx = 99999
    minindex = 0
    for i in range(0, len(c4)):
        temp = c4[i, 0] ** 2 + c4[i, 1] ** 2
        if minx > temp:
            minx = temp
            minindex = i
    for i in range(0, 4):
        c5[i] = c4[minindex]
        minindex = minindex + 1
        if minindex == 4:
            minindex = 0

    height = math.sqrt((c5[0, 0] - c5[1, 0]) ** 2 + (c5[0, 1] - c5[1, 1]) ** 2)
    height1 = math.sqrt((c5[2, 0] - c5[3, 0]) ** 2 + (c5[2, 1] - c5[3, 1]) ** 2)
    if height < height1:
        height = height1
    width = math.sqrt((c5[0, 0] - c5[3, 0]) ** 2 + (c5[0, 1] - c5[3, 1]) ** 2)
    width1 = math.sqrt((c5[2, 0] - c5[1, 0]) ** 2 + (c5[2, 1] - c5[1, 1]) ** 2)
    if width < width1:
        width = width1

    c2 = np.zeros((4, 2), dtype=np.float32)

    c2[1] = [0, height]
    c2[2] = [width, height]
    c2[3] = [width, 0]
    c2[0] = [0, 0]

    transmtx = cv2.getPerspectiveTransform(c5, c2)
    newim = np.zeros((np.int0(width), np.int0(height), 3), np.uint8)
    newim = cv2.warpPerspective(
        src, transmtx, (np.int0(width), np.int0(height)), newim)

    cv2.imshow("5", newim)
    cv2.imwrite("output6.jpg",newim)
    end = time.time()

    print(str(end - start))
    black = np.zeros(img.shape, np.uint8)
    for i in range(0, 4):
        cv2.circle(black, (int(c5[i, 0]), int(c5[i, 1])), 3, (128), 2)
        cv2.circle(black, (int(c2[i, 0]), int(c2[i, 1])), 3, (255), 2)
        cv2.putText(black, str(i), (int(c5[i, 0]), int(c5[i, 1])),
                    cv2.FONT_HERSHEY_TRIPLEX, 0.8, (128))
        cv2.putText(black, str(i), (int(c2[i, 0]), int(c2[i, 1])),
                    cv2.FONT_HERSHEY_TRIPLEX, 0.8, (255))


cv2.waitKey(0)
cv2.destroyAllWindows()

二、效果测试

部分校正样例：

1

2

3

4

结果：
1

2

3

4

总结

基本实现了矫正功能以及与后续代码的整合，测试效果时考虑了阴影与折叠等因素，经过调参后可以得到较理想的结果。
缺点与改进：1、只以四个角为基准透视变换可能导致后续单个方格识别准确率降低 2、准备对图片的像素和比例做一定的规范处理方便后续识别
后续工作：搭建以SpringBoot为框架的java后端