『科学计算』图像检测微型demo
这里是课上老师给出的一个示例程序,演示图像检测的过程,本来以为是传统的滑窗检测,但实际上引入了selectivesearch来选择候选窗,所以看思路应该是RCNN的范畴,蛮有意思的,由于老师的注释写的蛮好的,我基本就不画蛇添足了,这里记录下来,为加深理解cs231n的课程做个铺垫。,所以做个储备,实在不行还有开学不是么233
# coding: utf-8 #copyRight by heibanke #如需转载请注明出处 #<<用Python做深度学习1-数学基础>> #http://study.163.com/course/courseMain.htm?courseId=1050010 import numpy as np # 这里nnet是课程作业里实现的一个模块,参考资料里也会附上我的版本。大家也可以用自己做的版本。 from nnet.layers import FCLayer,Activation,SoftMaxCostLayer from nnet.neuralnetwork import neuralnetwork from nnet.helpers import one_hot # MNIST数据不再上传了,相信大家学到这里,这个数据应该都有一份,复制到文件夹内即可 import load_MNIST # 需要安装selectivesearch,pip install selectivesearch import selectivesearch from matplotlib import pyplot as plt import matplotlib.patches as mpatches # 需要安装opencv2 import cv2 %matplotlib inline
1. 用MNIST数据库训练分类器模型
这一步是我们之前课程里的重点,这里选用两层全连接神经网络模型进行训练。数据库的数据预测率能达到97%,大家可以根据自己喜好选择不同的模型试一下。
def get_model(): train_X,train_y,test_X,test_y = load_MNIST.get_data() n_classes = np.unique(train_y).size w_decay = 0.0001 nn = neuralnetwork( layers=[ FCLayer( n_out=128, weight_decay = w_decay, ), Activation('sigmoid'), FCLayer( n_out=n_classes, weight_decay = w_decay, ), Activation('softmax'), ], cost = SoftMaxCostLayer(), ) X = train_X.reshape(train_X.shape[0],28*28) Y_one_hot = one_hot(train_y) nn._setup(X, Y_one_hot) # Train neural network print('Training neural network') nn.train(X, train_y, learning_rate=1.0, max_epochs=8, batch_size=128) # Evaluate on training data error = nn.error(test_X.reshape(test_X.shape[0],28*28), test_y) print('Test error rate: %.4f' % error) return nn nn=get_model()
2.读入待测图片,并在待测试图片上用selective search算法获得物体窗口
待测图片是我自己在Photoshop上手写的数字,几个数字在一张图片上,不同大小,不同位置。
img = cv2.imread("test1.jpg") img_lbl, regions = selectivesearch.selective_search( img, scale=500, sigma=0.9, min_size=20) print regions[0] print len(regions)
{'labels': [0.0], 'rect': (0, 0, 511, 511), 'size': 243048} 49
# 接下来我们把窗口和图像打印出来,对它有个直观认识 fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6)) ax.imshow(img) for reg in regions: x, y, w, h = reg['rect'] rect = mpatches.Rectangle( (x, y), w, h, fill=False, edgecolor='red', linewidth=1) ax.add_patch(rect) plt.show()
3.定义规则来筛选窗口
candidates = [] for r in regions: # 重复的不要 if r['rect'] in candidates: continue # 太小和太大的不要 if r['size'] < 200 or r['size']>20000: continue x, y, w, h = r['rect'] # 太不方的不要 if w / h > 1.2 or h / w > 1.2: continue candidates.append((x,y,w,h)) print len(candidates) # 这一步的序号是事先设定好的,真正实现时不这样做,肯定会有多的窗口需要你以后来筛选。 candidates_re = [candidates[i] for i in [0,4,7,9,11]] print u"最终筛选后的窗口是:",candidates_re fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6)) ax.imshow(img) for x, y, w, h in candidates_re: rect = mpatches.Rectangle( (x, y), w, h, fill=False, edgecolor='red', linewidth=1) ax.add_patch(rect) plt.show()
最终筛选后的窗口是: [(47, 31, 65, 89), (335, 124, 84, 116), (127, 230, 65, 90), (343, 375, 41, 50), (183, 399, 73, 81)]
4.对窗口内图片进行处理,大小resize,转换灰度图,最终转换成为784的输入向量
img_sample = np.zeros((len(candidates_re),784)) i=0 for rect in candidates_re: x,y,w,h = rect if w>h: largewh = w else: largewh = h bord_size = int(largewh*0.2) img_cut = img[y-bord_size:y+largewh+bord_size,x-bord_size:x+largewh+bord_size,:] img_resize = cv2.resize(img_cut,(28,28),interpolation=cv2.INTER_NEAREST) gray=cv2.cvtColor(img_resize, cv2.COLOR_BGR2GRAY) img_sample[i,:]=gray.ravel() i+=1 # 把转换后的数据用图来显示 img_s=np.zeros((28,28*img_sample.shape[0])) for i in xrange(img_sample.shape[0]): img_s[:,i*28:28*(i+1)]=img_sample[i,:].reshape(28,28) fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6)) ax.imshow(img_s,cmap='gray') plt.show()
5.用训练好的模型对处理后的图片进行预测
label = nn.predict(img_sample/255) print u"每个窗口的预测值为:",label
每个窗口的预测值为: [8 5 3 5 0]
[注],检测失败了一个。