https://github.com/lucasb-eyer/pydensecrf/blob/master/examples/inference.py
1.運行
先運行看看實現的結果：
(deeplearning) userdeMBP:examples user$ python inference.py im1.png anno1.png out1.png Found a full-black pixel in annotation image, assuming it means 'unknown' label, and will thus not be present in the output! If 0 is an actual label for you, consider writing your own code, or simply giving your labels only non-zero values. 2 labels plus "unknown" 0: {0, 1, 2} Using generic 2D functions KL-divergence at 0: -543957.3854815669 KL-divergence at 1: -890605.7866870646 KL-divergence at 2: -919933.3682610085 KL-divergence at 3: -921683.1852052805 KL-divergence at 4: -922674.4361045817
im1.png和anno1.png是輸入圖片，out1.png為進行crf處理后的輸出圖片
im1.png和anno1.png為：
得到的輸出結果是：
可見效果變得很好
2.代碼分析
""" Adapted from the inference.py to demonstate the usage of the util functions. """ import sys import numpy as np import pydensecrf.densecrf as dcrf # Get im{read,write} from somewhere. try: from cv2 import imread, imwrite except ImportError: # Note that, sadly, skimage unconditionally import scipy and matplotlib, # so you'll need them if you don't have OpenCV. But you probably have them. from skimage.io import imread, imsave imwrite = imsave # TODO: Use scipy instead. from pydensecrf.utils import unary_from_labels, create_pairwise_bilateral, create_pairwise_gaussian if len(sys.argv) != 4: print("Usage: python {} IMAGE ANNO OUTPUT".format(sys.argv[0])) print("") print("IMAGE and ANNO are inputs and OUTPUT is where the result should be written.") print("If there's at least one single full-black pixel in ANNO, black is assumed to mean unknown.") sys.exit(1) fn_im = sys.argv[1]#輸入的圖片 print(fn_im) fn_anno = sys.argv[2]#輸入的圖片fn_im經過訓練后的網絡進行預測得到的結果 print(fn_anno) fn_output = sys.argv[3]#指定進行crf處理后的結果輸出 print(fn_output) ############################################################## ### Read images and annotation讀取輸入的兩個圖片fn_im和fn_anno### ############################################################## img = imread(fn_im) # Convert the annotation's RGB color to a single 32-bit integer color 0xBBGGRR #將fn_anno的三個uint8表示的RGB像素值放到一個uint32像素值中表示 #[0,7]位為R層的值，[8,15]為G層的值，[16,23]為B層的值 anno_rgb = imread(fn_anno).astype(np.uint32)#shape為(240, 320, 3) anno_lbl = anno_rgb[:,:,0] + (anno_rgb[:,:,1] << 8) + (anno_rgb[:,:,2] << 16)#shape變為了(240, 320) # Convert the 32bit integer color to 1, 2, ... labels. # Note that all-black, i.e. the value 0 for background will stay 0. # np.unique該函數是去除數組中的重復數字，並進行排序之后輸出 # 這就得到了整張圖中有的像素值序列 # #colors返回為[0,16384,4227072],說明圖片fn_anno只有這三種像素值 # labels的shape為(76800,)，其為anno_lbl中所有的像素值標上了對應的label # 在這里color=0時，對應的label為0;color=16384時，對應的label為1;color=4227072時，對應的label為2 # 黑色的像素值為0 colors, labels = np.unique(anno_lbl, return_inverse=True) # But remove the all-0 black, that won't exist in the MAP! # 移除像素值為0，即黑色的值 HAS_UNK = 0 in colors#若0存在於colors中，則HAS_UNK為True #在annotation圖像中的黑色像素，即color=0的像素，被假設為label='unknown'，不會在output中輸出 #如果0是一個對你來說有意義的label，那么更改你的代碼，或者盡量讓你的label為非0的數值 if HAS_UNK: print("Found a full-black pixel in annotation image, assuming it means 'unknown' label, and will thus not be present in the output!") print("If 0 is an actual label for you, consider writing your own code, or simply giving your labels only non-zero values.") colors = colors[1:]#然后將color=0從數組中移除 #else: # print("No single full-black pixel found in annotation image. Assuming there's no 'unknown' label!") # And create a mapping back from the labels to 32bit integer colors. # np.empty()返回一個隨機元素的矩陣,值類型為uint8，大小按照參數定義，這里 colorize = np.empty((len(colors), 3), np.uint8)#colorize.shape為(2,3) #下面將之前合並成0xBBGGRR格式的像素值又分成三層，得到各層像素值的值 colorize[:,0] = (colors & 0x0000FF)#得到R層的值，為[0,0], dtype=uint8 colorize[:,1] = (colors & 0x00FF00) >> 8#得到G層的值，為[ 64, 128], dtype=uint8 colorize[:,2] = (colors & 0xFF0000) >> 16#得到B層的值，[ 0, 64] # Compute the number of classes in the label image. # We subtract one because the number shouldn't include the value 0 which stands # for "unknown" or "unsure". # set(labels.flat)返回{0, 1, 2} # flat將數組變為一個迭代器，可以用for訪問數組每一個元素，可以使用索引labels.flat[0]來訪問第一個元素 # set(迭代對象) 函數創建一個無序不重復元素集，可進行關系測試，刪除重復數據 n_labels = len(set(labels.flat)) - int(HAS_UNK) #返回2，得到除去了label=0后還有兩個label print(n_labels, " labels", (" plus \"unknown\" 0: " if HAS_UNK else ""), set(labels.flat)) ########################### ### Setup the CRF model ### ########################### #上面處理完圖片fn_anno，得到labels和colors #接下來就是設置CRF模型了 use_2d = False #是否使用二維指定函數DenseCRF2D，這里設置為False，則說明使用的是一般函數DenseCRF # use_2d = True if use_2d: print("Using 2D specialized functions") # Example using the DenseCRF2D code d = dcrf.DenseCRF2D(img.shape[1], img.shape[0], n_labels) # get unary potentials (neg log probability) U = unary_from_labels(labels, n_labels, gt_prob=0.7, zero_unsure=HAS_UNK) d.setUnaryEnergy(U) # This adds the color-independent term, features are the locations only. # 創建顏色無關特征，這里只有位置特征，並添加到CRF中 d.addPairwiseGaussian(sxy=(3, 3), compat=3, kernel=dcrf.DIAG_KERNEL, normalization=dcrf.NORMALIZE_SYMMETRIC) # This adds the color-dependent term, i.e. features are (x,y,r,g,b). # 根據原始圖像img創建顏色相關特征和位置相關並添加到CRF中，特征為(x,y,r,g,b) d.addPairwiseBilateral(sxy=(80, 80), srgb=(13, 13, 13), rgbim=img, compat=10, kernel=dcrf.DIAG_KERNEL, normalization=dcrf.NORMALIZE_SYMMETRIC) else: print("Using generic 2D functions") # Example using the DenseCRF class and the util functions # 使用DenseCRF類和util函數 # n_labels為2，從上面對fn_anno的分析可知有兩個label d = dcrf.DenseCRF(img.shape[1] * img.shape[0], n_labels) # get unary potentials (neg log probability) # 得到一元勢(即去負對數)，labels為對所有像素值標注label后的數組，label類型n_labels=2， U = unary_from_labels(labels, n_labels, gt_prob=0.7, zero_unsure=HAS_UNK) #U.shape為(2, 76800),即(n_labels,len(labels)) d.setUnaryEnergy(U) #將一元勢添加到CRF中 # This creates the color-independent features and then add them to the CRF # 創建顏色無關特征，這里只有位置特征，並添加到CRF中 feats = create_pairwise_gaussian(sdims=(3, 3), shape=img.shape[:2]) #shape為(240, 320) d.addPairwiseEnergy(feats, compat=3, kernel=dcrf.DIAG_KERNEL, normalization=dcrf.NORMALIZE_SYMMETRIC) # This creates the color-dependent features and then add them to the CRF # 根據原始圖像img創建顏色相關和位置相關特征並添加到CRF中，特征為(x,y,r,g,b) feats = create_pairwise_bilateral(sdims=(80, 80), schan=(13, 13, 13), img=img, chdim=2) d.addPairwiseEnergy(feats, compat=10, kernel=dcrf.DIAG_KERNEL, normalization=dcrf.NORMALIZE_SYMMETRIC) #################################### ### Do inference and compute MAP ### #################################### #上面就將相應的CRF構建好了 #然后要做的就是對img根據fn_anno得到的label和colors結果進行CRF推理 #然后得到輸出值fn_output了 # Run five inference steps.迭代5次 Q = d.inference(5) # Find out the most probable class for each pixel. # 找出每個像素最可能的類 # np.argmax取出Q元素中最大的值對應的索引,axis=0按列查找 MAP = np.argmax(Q, axis=0) # MAP,MAP.shape返回 # (array([1, 1, 1, ..., 1, 1, 1]), (76800,)) # 將MAP(標簽)轉換回相應的顏色並保存圖像。 #注意，這里不再有“unknown”標簽，不管我們一開始擁有什么。 #colorize返回兩個label的color[16384,4227072]對應的RGB的值 #16384對應[ 0, 64, 0]，4227072對應[ 0, 128, 64] #array([[ 0, 64, 0], # [ 0, 128, 64]], dtype=uint8) #MAP中1值對應的是4227072即[ 0, 128, 64] MAP = colorize[MAP,:] #MAP.shape為(76800, 3)，這就是最后的結果 #將MAP轉成img相同的大小，就能夠得到最后的結果了 imwrite(fn_output, MAP.reshape(img.shape)) # Just randomly manually run inference iterations # 這里是手動實現迭代推理 Q, tmp1, tmp2 = d.startInference() for i in range(5): print("KL-divergence at {}: {}".format(i, d.klDivergence(Q))) d.stepInference(Q, tmp1, tmp2)