图像识别快速实现

发布时间:2024年01月03日

文本的跑通了,接下来玩玩图片场景

1. 引入模型

再另起类test_qdrant_img.py,转化图片用到的模型和文本不太一样,我们这里使用ResNet-50模型

import unittest
from qdrant_client.http.models import Distance, VectorParams
from qdrant_client import QdrantClient
import torch
import torchvision.transforms as transforms
from PIL import Image

class TestQDrantImg(unittest.TestCase):

    def setUp(self):
        self.collection_name = "img_collection"
        self.client = QdrantClient("localhost", port=6333)
        # 加载ResNet-50模型
        self.model = torch.hub.load('pytorch/vision:v0.10.0', 'resnet50', pretrained=True)
        self.model.eval()
        # 图像预处理
        self.preprocess = transforms.Compose([
            # 图像调整为256*256
            transforms.Resize(256), 
            # 中心裁剪为224*224
            transforms.CenterCrop(224), 
            # 转换为张量,像素值从范围[0,255]缩放到范围[0,1],RGB(红绿蓝)转换为通道顺序(即 RGB 顺序)
            transforms.ToTensor(), 
            # 应用归一化,减去均值(mean)并除以标准差(std)
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])

2. 添加图片向量

我们先创建一个新集合

def test_create_collection(self):
    self.client.create_collection(
        collection_name=self.collection_name,
        vectors_config=VectorParams(size=1000, distance=Distance.EUCLID),
    )

往集合里分别添加1个猫的图片和1个狗的图片

    def test_img_vector(self):
        # 加载并预处理图像
        id = 1
        image_path = './img/cat1.png'
        # id = 2
        # image_path = './img/dog1.png'
        image = Image.open(image_path)
        image_tensor = self.preprocess(image)
        # 在第0维度上添加一个维度,将图像张量转换为形状为 (1, C, H, W) 的张量,其中 C 是通道数,H 是高度,W 是宽度
        image_tensor = torch.unsqueeze(image_tensor, 0)
        with torch.no_grad():
            # 去除维度为1的维度,将特征向量的形状从 (1, D) 转换为 (D,)
            feature_vector = self.model(image_tensor).squeeze().tolist()

        operation_info = self.client.upsert(
            collection_name=self.collection_name,
            points=[{'id': id, 'vector': feature_vector, 'payload': {"image_path": image_path}}]
        )

        print(operation_info)

3. 匹配图片向量

然后用其他猫狗的图片来做搜索匹配

    def test_search(self):
        # 加载并预处理图像
        image_path = './img/cat2.png'
        # image_path = './img/dog2.png'
        # image_path = './img/cat3.png'
        image = Image.open(image_path)
        image_tensor = self.preprocess(image)
        image_tensor = torch.unsqueeze(image_tensor, 0)
        with torch.no_grad():
            feature_vector = self.model(image_tensor).squeeze().tolist()

        search_result = self.client.search(
            collection_name=self.collection_name, query_vector=feature_vector, limit=3
            , with_vectors=True, with_payload=True
        )

        print(search_result)

结果:

[ScoredPoint(id = 1, version = 0, score = 68.21013, payload = {

'image_path': './img/cat1.png'

}, vector = [...]),

ScoredPoint(id = 2, version = 1, score = 85.10757, payload = {

'image_path': './img/dog1.png'

}, vector = [...])]

当使用猫2猫3作为查询条件时,跟猫1记录的score(向量距离)较小;

同理,使用狗2作为查询条件时,跟狗1记录的score(向量距离)较小

文章来源:https://blog.csdn.net/cxs812760493/article/details/135346484
本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。