非极大值抑制(nms)算法功能及python实现

    有什么比一张图更能说明问题呢。nms广泛应用于边缘检测，人脸检测和目标检测等，用于消除冗余的框。也有专门研究它的论文。如下就Faster-RCNN_TF中nms的python源码进行注释。

图1 nms的一个功能示意图，图片来源

• 源码文件nms.py

def nms(dets, thresh):
    x1 = dets[:, 0]
    y1 = dets[:, 1]
    x2 = dets[:, 2]
    y2 = dets[:, 3]
    scores = dets[:, 4]

    areas = (x2 - x1 + 1) * (y2 - y1 + 1) #所有box面积
    print "all box aress: ", areas
    order = scores.argsort()[::-1] #降序排列得到scores的坐标索引

    keep = []
    while order.size > 0:
        i = order[0] #最大得分box的坐标索引
        keep.append(i)
        xx1 = np.maximum(x1[i], x1[order[1:]]) 
        yy1 = np.maximum(y1[i], y1[order[1:]])
        xx2 = np.minimum(x2[i], x2[order[1:]])
        yy2 = np.minimum(y2[i], y2[order[1:]]) #最高得分的boax与其他box的公共部分(交集)

        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1) #求高和宽，并使数值合法化
        inter = w * h #其他所有box的面积
        ovr = inter / (areas[i] + areas[order[1:]] - inter)  #IOU:交并比

        inds = np.where(ovr <= thresh)[0] #ovr小表示两个box交集少，可能是另一个物体的框，故需要保留
        order = order[inds + 1]  #iou小于阈值的框

    return keep

• 测试用例

import cv2
import numpy as np
import random

img=np.zeros((300,400), np.uint8)
dets=np.array([[83,54,165,163,0.8], [67,48,118,132,0.5], [91,38,192,171,0.6]], np.float)

img_cp=img.copy()
for box in dets.tolist():	#显示待测试框及置信度
    x1,y1,x2,y2,score=int(box[0]),int(box[1]),int(box[2]),int(box[3]),box[-1]
    y_text=int(random.uniform(y1, y2))
    cv2.rectangle(img_cp, (x1,y1), (x2, y2), (255, 255, 255), 2)
    cv2.putText(img_cp, str(score), (x2-30, y_text), 2,1, (255, 255, 0))
cv2.imshow("ori_img", img_cp)

rtn_box=nms(dets, 0.3)	#0.3为faster-rcnn中配置文件的默认值
cls_dets=dets[rtn_box, :]
print "nms box:", cls_dets

img_cp=img.copy()
for box in cls_dets.tolist():
    x1,y1,x2,y2,score=int(box[0]),int(box[1]),int(box[2]),int(box[3]),box[-1]
    y_text=int(random.uniform(y1, y2))
    cv2.rectangle(img_cp, (x1,y1), (x2, y2), (255, 255, 255), 2)
    cv2.putText(img_cp, str(score), (x2-30, y_text), 2,1, (255, 255, 0))

    输出图像为

参考文献：

1. 更详尽的文章参看https://www.pyimagesearch.com/2014/11/17/non-maximum-suppression-object-detection-python/