mxnet根据相似度进行人脸样本对图片清理

浅浅的花香味﹌ 2022-10-09 12:13 48阅读 0赞

清理的角度：

找出 ID (gallery) 和它的 prob 最不相似 的 数据 （可能是标注错误，不同ID的数据错误混杂等）  
找出 ID 和别的 prob 最相似 的 数据 （可能是标注错误，数据重复拷贝，同一个ID多次采集等）  
找出ID与ID之间最相似的。也是为了排除重复数据或者重复采集  
下面的方法主要关注角度1、角度3 ，实际中 角度2 也需要非常关注！

角度1  
背景：每个ID有一张ID图片和多张prob图片，id图片命名例如：AA20210501S0338.jpg prob图片命名例如：AA20210501S0338\_m001.jpg  
做法：加载模型，提取特征，计算id和prob的平均相似度，排序，保存

import numpy as np  
import os  
import mxnet as mx  
from menpo.visualize import print\_progress  
from collections import namedtuple

Batch = namedtuple('Batch', \['data'\])

def constructmodel(prefix, ctx, epoch):  
  sym, arg\_params, aux\_params = mx.model.load\_checkpoint(prefix, epoch)  
  all\_layers = sym.get\_internals()  
  sym = all\_layers\['fc1\_output'\]  
  model = mx.mod.Module(symbol=sym, context=ctx, label\_names = None)  
  model.bind(for\_training=False, data\_shapes=\[('data', (1, 3, 112, 112))\])  
  model.set\_params(arg\_params, aux\_params)  
  return model

def extractfeature(model, img, featurelength = 512):  
  model.forward(Batch(\[img\]))  
  net\_out = model.get\_outputs()  
  emb = net\_out\[0\].asnumpy()  
  emb = emb.reshape((featurelength, 1))  
  return emb

def id\_getfeatures\_dict(model, imglist, image\_root="",featurelength = 512):  
  features = \[\]  
  feature\_dict = \{\}  
  with open(imglist,'r') as file:  
    lines = file.readlines()  
    for line in print\_progress(lines):  
        vec = line.strip().split('\\t')       
        imagepath = vec\[0\]  
        key = os.path.splitext(imagepath)\[0\]  
        ori\_img = mx.image.imread(os.path.join(image\_root,imagepath))  
        oriimg = ori\_img.astype('float32')  
        oriimg = mx.image.imresize(oriimg, 112, 112)  \# resize  
        oriimg = oriimg.transpose((2, 0, 1))  \# Channel first  
        oriimg = oriimg.expand\_dims(axis=0)  \# batchify  
        emb = extractfeature(model, oriimg)  
        \# feature\_dict\[imagepath\] = emb  
        feature\_dict.setdefault(key, \[\]).append(emb)  
        \# features.append(emb)  
  return feature\_dict

def prob\_getfeatures\_dict(model, imglist, image\_root="",featurelength = 512):  
  features = \[\]  
  feature\_dict = \{\}  
  with open(imglist,'r') as file:  
    lines = file.readlines()  
    for line in print\_progress(lines):  
        vec = line.strip().split('\\t')  
        imagepath = vec\[0\]  
        key = imagepath.split('\_m')\[0\]  
        ori\_img = mx.image.imread(os.path.join(image\_root,imagepath))  
        oriimg = ori\_img.astype('float32')  
        oriimg = mx.image.imresize(oriimg, 112, 112)  \# resize  
        oriimg = oriimg.transpose((2, 0, 1))  \# Channel first  
        oriimg = oriimg.expand\_dims(axis=0)  \# batchify  
        emb = extractfeature(model, oriimg)  
        \# feature\_dict\[imagepath\] = emb  
        feature\_dict.setdefault(key, \[\]).append(emb)  
        \# features.append(emb)  
  return feature\_dict

def cal\_sim(id\_fdict,prob\_fdict):  
    assert len(id\_fdict.keys()) == len(prob\_fdict.keys())  
    sim\_dict = \{\}  
    sim\_num\_dict = \{\}  
    f = open('../id\_prob\_sims.txt', 'w')  
    for k in print\_progress(id\_fdict.keys()):  
        idfea = id\_fdict\[k\]  
        sims = \[\]  
        for probfea in prob\_fdict\[k\]:  
            s = np.dot(np.squeeze(idfea), np.squeeze(probfea)) / (np.linalg.norm(idfea) \* np.linalg.norm(probfea))  
            sims.append(s)  
        sim\_dict\[k\] = sum(sims)/len(sims)  
        sim\_num\_dict\[k\] = len(sims)  
    sim\_tuple = sorted(sim\_dict.items(), key=lambda item:item\[1\])  
    for k in sim\_tuple:  
        f.write('%s\\t%s\\t%s\\n'%(k\[0\],sim\_num\_dict\[k\[0\]\],k\[1\]))  
    f.close()

modelPth = '/home/model,368'  
prefix,epoch=modelPth.split(',')  
ctx=\[mx.gpu(0)\]        
featurelen = 512  
id\_img\_list = '../clean\_id\_list.txt'  
prob\_img\_list = '../clean\_prob\_list.txt'  
ID\_path = '../ID\_img\_aligned'  
prob\_path = '../prob\_img\_aligned\_clean'  
model = constructmodel(prefix, ctx, int(epoch))  
id\_features\_dict = id\_getfeatures\_dict(model, id\_img\_list, ID\_path, featurelen)  
prob\_features\_dict = prob\_getfeatures\_dict(model, prob\_img\_list, prob\_path, featurelen)  
cal\_sim(id\_features\_dict,prob\_features\_dict)  
  
注意：需要选用性能比较强的模型进行特征提取。  
相似度比较低的ID需要人工check 其ID图片和prob图片的一致性

角度3  
\# 其他函数相同

import itertools

def cal\_sim(id\_fdict,prob\_fdict):  
    assert len(id\_fdict.keys()) == len(prob\_fdict.keys()),'%s-%s'%(len(id\_fdict.keys()),len(prob\_fdict.keys()))  
    sim\_dict = \{\}  
    sim\_num\_dict = \{\}  
    f = open('../id\_inter\_sims.txt', 'w')  
    id\_combins = list(itertools.combinations(id\_fdict.keys(), 2))  
    for (idk1,idk2) in print\_progress(id\_combins):  
        idfea = id\_fdict\[idk1\]  
        probfea = id\_fdict\[idk2\]  
        sims = \[\]  
        s = np.dot(np.squeeze(idfea), np.squeeze(probfea)) / (np.linalg.norm(idfea) \* np.linalg.norm(probfea))  
        sim\_dict\['%s\\t%s'%(idk1,idk2)\] = s  
    sim\_tuple = sorted(sim\_dict.items(), key=lambda item:item\[1\])  
    for k in sim\_tuple:  
        k0,k1 = k\[0\].split('\\t')  
        f.write('%s\\t%s\\t%s\\n'%(k0,k1,k\[1\]))  
    f.close()

modelPth = '/home/model,368'  
prefix,epoch=modelPth.split(',')  
ctx=\[mx.gpu(0)\]  
featurelen = 512  
id\_img\_list = '../clean\_id\_list.txt'  
ID\_path = '../ID\_img\_aligned'  
model = constructmodel(prefix, ctx, int(epoch))  
id\_features\_dict = id\_getfeatures\_dict(model, id\_img\_list, ID\_path, featurelen)  
cal\_sim(id\_features\_dict,id\_features\_dict)  
  
原文链接：https://blog.csdn.net/qxqxqzzz/article/details/118091009