NMS——非极大值抑制

非极大值抑制（NMS）：去除掉与得分最高的目标框IOU值>thresh的目标框，保留与得分最高的目标框IOU值<thresh的目标框；

算法：

输入：dets[x1,y1,x2,y2,score]为框列表 ,thresh为IOU值
1.将所有框按score值从大到小排序，将索引存储为列表order
2.取order中的第一个值（对应为score最大的框）
3.计算该框与其他框的IOU值
4.选取IOU<thresh的框  
5.将选定的框的索引赋值给order
6.如果order中有值，则执行步骤2

下面时Fast R-CNN关于NMS的源代码(python版)，Faster R-CNN也是用的这段代码。

import numpy as np
import cv2
def py_cpu_nms(dets, thresh):
    """Pure Python NMS baseline."""
    x1 = dets[:, 0]
    y1 = dets[:, 1]
    x2 = dets[:, 2]
    y2 = dets[:, 3]
    scores = dets[:, 4]
    areas = (x2 - x1 + 1) * (y2 - y1 + 1)
    #数组值从大到小的索引值，argsort()得到数组值从小到大的索引值，[::-1]倒序值
    order = scores.argsort()[::-1]
    keep = []
    while order.size > 0:
        i = order[0]
        keep.append(i)
        #求其他框与评分最高的框的共有区域
        xx1 = np.maximum(x1[i], x1[order[1:]])  #xx1的值是按order中的排序返回的
        yy1 = np.maximum(y1[i], y1[order[1:]])
        xx2 = np.minimum(x2[i], x2[order[1:]])
        yy2 = np.minimum(y2[i], y2[order[1:]])
        w = np.maximum(0.0, xx2 - xx1 + 1)
        h = np.maximum(0.0, yy2 - yy1 + 1)
        inter = w * h
        #共有区域与两个框总面积的比值，获取其他目标框与得分最高的框IOU值
        ovr = inter / (areas[i] + areas[order[1:]] - inter)
        inds = np.where(ovr <= thresh)[0]
        order = order[inds + 1] #获取与最高得分的边框IOU值小于thresh的框
    return keep

图像显示

def score(dets,y):
    #物体框的score=物体框与目标框的IOU值/目标框的
    myscore=[]
    ovr = ((y[2] - y[0] + 1) * (y[3] - y[1] + 1))
    for i in range(len(dets)):
        x1=np.maximum(dets[i][0],y[0])
        y1 = np.maximum(dets[i][1], y[1])
        x2 = np.minimum(dets[i][2], y[2])
        y2 = np.minimum(dets[i][3], y[3])
        w=np.maximum(0.0,x2-x1+1)
        h = np.maximum(0.0, y2 - y1 + 1)
        inter=w*h
        myscore.append(inter/ovr)
    return myscore

if __name__ == '__main__':
    dets = np.array([[212., 312., 812., 712.],
                     [262., 262., 762., 762.],
                     [312., 213., 712., 812.],
                     ])
    y = np.array([542., 400., 842., 700.])
    score_1=np.array([score(dets,y)]).T        #形状为(3, 1),[0.90033223,0.73421927,0.56810631]]
    dets=np.hstack((dets,score_1))             #形状为(3, 5)
    keep=py_cpu_nms(dets,0.2)                  #keep=[0],即得分最高的框为dets[0]

    # 图像只支持uint8类型数据，可以通过astype转变数据类型
    img = np.array(np.random.uniform(0, 255, size=(1024, 1024, 3)), dtype=np.uint8)
    # print(img)
    # img=cv2.imread(path)
    print(type(img))
    w, h, c = img.shape

    for i in range(3):
        x1 = int(dets[i][0])
        y1 = int(dets[i][1])
        x2 = int(dets[i][2])
        y2 = int(dets[i][3])
        cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 3)
        cv2.putText(img, str(i), (x1, y1), cv2.FONT_HERSHEY_COMPLEX_SMALL, 4, (0, 255, 0), 3)

    cv2.rectangle(img, (int(y[0]), int(y[1])), (int(y[2]), int(y[3])), (0, 0, 255), 3)
    cv2.rectangle(img, (int(dets[keep][0]), int(dets[keep][1])), (int(dets[keep][2]), int(dets[keep][3])), (255, 0, 255), 3)
    cv2.imshow('img', img)
    cv2.waitKey()
    cv2.destroyAllWindows()