项目实训（十一）—相似场景聚类

前言：

上一篇博客中已经将所有的场景视频的特征提取出来了，下面介绍一下如何将相似的场景视频聚类到一起。

代码实现：

1、读取h5特征文件：

def H5Filepocess():
    h5_file_path_dir = '/opt/data/private/xuyunyang/EasyCut/' + args.ID + '/' + args.ID_VideoName + '/SceneFeature'
    dirs = os.listdir(h5_file_path_dir)
    for file in dirs:
        h5_file_path = h5_file_path_dir + '/' + file
        h5 = h5py.File(h5_file_path, 'r')
        data = h5['video_1']['features'][...]
        h5FileName_feature[h5_file_path] = data

2、相似度对比：
思路是对于每一个视频，计算与除自己外的所有其他视频的相似度，相似度计算使用余弦相似度方法，设定相似度阈值为0.95，大于阈值的视频就被添加到列表中，该列表之后作为字典结构的值存在，键则为对应的目标视频路径。

def CalculateDistance(path):
    distance = dict()
    A = h5FileName_feature[path]
    a = A
    a = np.average(A, axis=0)  # 按列求均值
    for i in h5FileName_feature.keys():
        if i == path:
            continue

        B = h5FileName_feature[i]
        b = B
        b = np.average(B, axis=0)
        # cosin相似度(余弦相似度)方法:把图片表示成一个向量，通过计算向量之间的余弦距离来表征两张图片的相似度。
        a_norm = np.linalg.norm(a)
        b_norm = np.linalg.norm(b)
        cosin = np.dot(a, b) / (a_norm * b_norm)
        distance[i.split('.')[0]+'.mp4'] = cosin

    # distance=sorted(distance.items(),key=lambda x:x[1],reverse=True) #从大到小排序
    # 筛选出相似度大于0.95的
    distance1 = {k: v for k, v in distance.items() if v >= 0.95}
    result_list = list(distance1.keys())

    return result_list

3、以字典结构存储结果：

def generate_result():
    H5Filepocess()
    # print(h5FileName_feature)
    for key in h5FileName_feature.keys():
        result_dict[key.split('.')[0]+'.mp4'] = CalculateDistance(key)
        # print('key:',key)
        # print('reasult_dic_list:',result_dict[key])

4、将结果保存为json文件：

result_json = json.dumps(result_dict)
    with open(
            '/opt/data/private/xuyunyang/EasyCut/' + args.ID + '/' + args.ID_VideoName + '/SceneSimilar/' + args.ID_VideoName+'_scene_clustering.json',
            'w') as f:
        f.write(result_json)
    print('END ALL')

完整代码如下：

import json
import numpy as np
import os
import h5py
import argparse

parser = argparse.ArgumentParser("Compute scene similarity of videos and generate a clusterd result")
parser.add_argument('--ID', type=str)
parser.add_argument('--ID_VideoName', type=str)
args = parser.parse_args()

h5FileName_feature = dict()
result_dict = dict()


def H5Filepocess():
    h5_file_path_dir = '/opt/data/private/xuyunyang/EasyCut/' + args.ID + '/' + args.ID_VideoName + '/SceneFeature'
    dirs = os.listdir(h5_file_path_dir)
    for file in dirs:
        h5_file_path = h5_file_path_dir + '/' + file
        h5 = h5py.File(h5_file_path, 'r')
        data = h5['video_1']['features'][...]
        h5FileName_feature[h5_file_path] = data


def CalculateDistance(path):
    distance = dict()
    A = h5FileName_feature[path]
    a = A
    a = np.average(A, axis=0)  # 按列求均值
    for i in h5FileName_feature.keys():
        if i == path:
            continue

        B = h5FileName_feature[i]
        b = B
        b = np.average(B, axis=0)
        # cosin相似度(余弦相似度)方法:把图片表示成一个向量，通过计算向量之间的余弦距离来表征两张图片的相似度。
        a_norm = np.linalg.norm(a)
        b_norm = np.linalg.norm(b)
        cosin = np.dot(a, b) / (a_norm * b_norm)
        distance[i.split('.')[0]+'.mp4'] = cosin

    # distance=sorted(distance.items(),key=lambda x:x[1],reverse=True) #从大到小排序
    # 筛选出相似度大于0.95的
    distance1 = {k: v for k, v in distance.items() if v >= 0.95}
    result_list = list(distance1.keys())

    return result_list


def generate_result():
    H5Filepocess()
    # print(h5FileName_feature)
    for key in h5FileName_feature.keys():
        result_dict[key.split('.')[0]+'.mp4'] = CalculateDistance(key)
        # print('key:',key)
        # print('reasult_dic_list:',result_dict[key])


if __name__ == '__main__':
    generate_result()
    result_json = json.dumps(result_dict)
    with open(
            '/opt/data/private/xuyunyang/EasyCut/' + args.ID + '/' + args.ID_VideoName + '/SceneSimilar/' + args.ID_VideoName+'_scene_clustering.json',
            'w') as f:
        f.write(result_json)
    print('END ALL')