目录
实验目的及实验内容 1
实验目的： 1
实验内容： 1
原理分析： 1
实验环境 13
实验步骤及实验过程分析 13
实验结果总结 18
实验目的及实验内容
（本次实验所涉及并要求掌握的知识；实验内容；必要的原理分析）
实验目的：
使用 python 进行图像处理
实验内容：
自己找一张图片，用 OpenCV 的仿射变换实现图片缩放。
理解 HOG、ORC 过程，修改 digits.py 或独立编程，实现数字图像的识别，要求分别使用 SVM、knn（第六次课）以及 cnn 神经网络模型（第七次课），以群文件中 digit_data.rar 中 train 文件夹下数据作为训练集，test 文件夹下图片作为测试集，并计算测试集上分类正确率，本文转载自http://www.biyezuopin.vip/onews.asp?id=16706并自己找测试数据（可以是一个或多个数字）。
实验环境
（本次实验所使用的器件、仪器设备等的情况）
处理器：Intel® Core™ i5-9300H CPU @ 2.40GHz 2.40 GHz (2)操作系统环境：Windows 10 家庭中文版 x64 19042.867
编程语言：Python 3.8
其他环境：16 GB 运行内存
IDE 及包管理器：JetBrains PyCharm .1 x64， anaconda 3 for Windows（conda 4.9.0）
实验步骤及实验过程分析
（详细记录实验过程中发生的故障和问题，进行故障分析，说明故障排除的过程及方法。根据具体实验，记录、整理相应的数据表格、绘制曲线、波形等）
说明：
本篇实验报告所记录的内容仅为写报告时（/05/12）的情况，可能与实际实验时/04/28）结果有出入。
一切以实际运行时所得到的结果为准。
用 OpenCV 的仿射变换实现图片缩放：
import numpy as npimport torchfrom torch import nnimport torch.nn.functional as Fimport picklefrom utils import label_2_id, id_2_labelfrom torch.utils.data import TensorDataset, DataLoaderimport matplotlib.pyplot as pltimport cv2 as cvimport osclass CnnImage(object):def __init__(self, data_dir='../digit_data', save_dir='./data', batch_size=64,epochs=50, learning_rate=0.0005, use_gpu=False):self.data_dir = data_dirself.mode = 'train'self.save_dir = save_dirself.batch_size = batch_sizeif not use_gpu:self.device = torch.device("cpu")print('use device: cpu')else:self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")if self.device == torch.device("cpu"):print('no usable cuda detected! Use cpu instead')else:print('use device: gpu')self.epochs = epochsself.lr = learning_rateself.loss_ = []if not os.path.exists(self.save_dir):os.mkdir(self.save_dir)self.model = nn.Sequential(nn.Conv2d(1, 25, 3, 1),nn.MaxPool2d(2, 2),nn.Conv2d(25, 50, 3, 1),nn.MaxPool2d(2, 2),nn.Flatten(),nn.Linear(50 * 5 * 5, 10))self.model.to(self.device)self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.lr)def train(self, train_loader, test_loader, save=True):best_test = Nonetmp_loss = []best_model = self.modelfor epoch in range(self.epochs):for batch, label in train_loader:y_predict = self.model(batch)loss = F.cross_entropy(y_predict, label)tmp_loss.append(loss.item())loss.backward()self.optimizer.step()self.optimizer.zero_grad()self.loss_.append(sum(tmp_loss)/len(tmp_loss))correct_percent = 0error_num = 0with torch.no_grad():for batch, label in test_loader:y_predict = self.model(batch)top_p, top_class = torch.topk(y_predict, 1)top_class = top_class.squeeze(1)correct_percent += torch.sum(label == top_class).item()error_num += torch.sum(label != top_class).item()correct_percent /= len(test_loader) * self.batch_sizeif best_test is None or best_test < correct_percent:best_test = correct_percentbest_model = self.modelself.model = best_modelif save:tar = self.save_dir + '/' + 'model_cnn.pkl'with open(tar, 'wb') as f:pickle.dump(self.model, f)print('model saved successfully, see it in: ', tar)print('训练出的最好的模型在测试集上的正确率: ' + str(best_test))def load_model(self):tar = self.save_dir + '/' + 'model_cnn.pkl'with open(tar, 'rb') as f:self.model = pickle.load(f)print('model loaded successfully!')def predict(self, path):if os.path.isdir(path):files = os.listdir(path)else:print('no files detected!')return Nonex = []y = []for item in files:file = path + '/' + itemimg = cv.imread(file, 0) # 读灰度图re_shape = cv.resize(img, (28, 28), interpolation=cv.INTER_CUBIC)# shape: (28, 28), size of the reshaped imagex.append(re_shape)y.append(0)x_set = np.array(x)# array, shape:(n, 28, 28)y_set = np.array(y)# array, shape:(n,)x_tens = torch.tensor(x_set, dtype=torch.float).unsqueeze(1).to(self.device)y_tens = torch.tensor(y_set, dtype=torch.long).to(self.device)# x_tens: torch.Tensor, torch.Size([n, 1, 28, 28])# y_tens: torch.Tensor, torch.Size([n])sets = TensorDataset(x_tens, y_tens)loader = DataLoader(sets, self.batch_size, True)with torch.no_grad():for batch, label in loader:y_predict = self.model(batch)top_p, top_class = torch.topk(y_predict, 1)top_class = top_class.squeeze(1)return top_class.cpu().tolist() # 返回的结果在CPU上， 是个列表def set_data(self, save=True):sets = self.save_dir + '/sets.pkl'if os.path.exists(sets):with open(sets, 'rb') as f:loader = pickle.load(f)train_loader = loader[0]test_loader = loader[1]print('dataset loaded successfully!')else:self.mode = 'train'train_dataset = self.dataset()train_loader = DataLoader(train_dataset, self.batch_size, True)self.mode = 'test'test_dataset = self.dataset()test_loader = DataLoader(test_dataset, self.batch_size, True)print('dataset built successfully!')if save:loader = (train_loader, test_loader)with open(sets, 'wb') as f:pickle.dump(loader, f)print('dataset saved in: ', sets)return train_loader, test_loaderdef dataset(self):path = self.data_dir + '/' + self.modelabels = os.listdir(path)x = []y = []for label in labels:img_path = path + '/' + labelfiles = os.listdir(img_path)for item in files:file = img_path + '/' + itemimg = cv.imread(file, 0) # 读灰度图# shape: (28, 28), size of the imagex.append(img)y.append(label_2_id[label])x_set = np.array(x)# array, shape:(n, 28, 28)x_tens = torch.tensor(x_set, dtype=torch.float).unsqueeze(1).to(self.device)# torch.Tensor, torch.Size([n, 1, 28, 28])y_set = np.array(y)# array, shape:(n,)y_tens = torch.tensor(y_set, dtype=torch.long).to(self.device)# torch.Tensor, torch.Size([n])sets = TensorDataset(x_tens, y_tens)return setsdef train_():epochs = 100cnn = CnnImage(data_dir='../digit_data', save_dir='./data', batch_size=64,use_gpu=True, epochs=epochs, learning_rate=0.0005)train_loader, test_loader = cnn.set_data(save=True)cnn.train(train_loader=train_loader, test_loader=test_loader, save=True)loss = cnn.loss_plt.rcParams['font.sans-serif'] = ['SimHei']plt.rcParams['axes.unicode_minus'] = Falsefig = plt.figure()ax = fig.add_subplot(1, 1, 1)ax.set_title('epoch - loss')ax.set_xlabel('epoch')ax.set_ylabel('loss')x_axis = [i for i in range(len(loss))]plt.plot(x_axis, loss)plt.show()def test():cnn = CnnImage(data_dir='../digit_data', save_dir='./data', batch_size=64,use_gpu=True)cnn.load_model()test_path = 'testimg'pre = cnn.predict(test_path)if pre is None:returnelse:files = os.listdir(test_path)for i in range(len(files)):res = files[i] + ' 的预测结果为: ' + str(pre[i])print(res)if __name__ == '__main__':train_()test()
