2020-10-10caffe网络推理图像数据加载中的坑

caffe网络推理图像数据加载中的坑（Python切片）

1.人脸检测MobileNetSSD网络推理demo（正确的）

// An highlighted block
import numpy as np  
import sys,os  
import cv2
import caffe  
net_file= 'deploy.prototxt'
caffe_model='test.caffemodel'
test_dir = "./rgb_image"

if not os.path.exists(caffe_model):
    print("use merge_bn.py to generate it.")
    exit()
net = caffe.Net(net_file,caffe_model,caffe.TEST)  

CLASSES = ('background','face')
def preprocess_mssd(src):
    img = cv2.resize(src, (256,256))
    image = img.copy()
    print(img.shape)
    print(img.dtype)
    # img[:,:,0] = img[:,:,0] - 127.5
    # img[:,:,1] = img[:,:,1] - 127.5
    # img[:,:,2] = img[:,:,2] - 127.5
    img = img -127.5
    img = img * (1.0/127.5)
    return img, image

def postprocess(img, out):   
    h = img.shape[0]
    w = img.shape[1]
    box = out['detection_out'][0,0,:,3:7] * np.array([w, h, w, h])

    cls = out['detection_out'][0,0,:,1]
    conf = out['detection_out'][0,0,:,2]
    return (box.astype(np.int32), conf, cls)

def detect(imgfile):
    origimg = cv2.imread(imgfile)
    img, origimg = preprocess_mssd(origimg)

    img = img.astype(np.float32)
    img = img.transpose((2, 0, 1))

    net.blobs['data'].data[...] = img
    out = net.forward()  
    box, conf, cls = postprocess(origimg, out)
    print("imagefile: ",imgfile)
    for i in range(len(box)):
       p1 = (box[i][0], box[i][1])
       p2 = (box[i][2], box[i][3])
       print("(",box[i][0],",", box[i][1],")",",","(",box[i][2],",", box[i][3],")")
       cv2.rectangle(origimg, p1, p2, (0,255,0))
       p3 = (max(p1[0], 15), max(p1[1], 15))
       title = "%s:%.2f" % (CLASSES[int(cls[i])], conf[i])
       cv2.putText(origimg, title, p3, cv2.FONT_ITALIC, 0.6, (0, 255, 0), 1)
    cv2.imshow("ssd", origimg)
 
    k = cv2.waitKey(0) & 0xff
        #Exit if ESC pressed
    if k == 27 : return False
    return True

for f in os.listdir(test_dir):
    if detect(test_dir + "/" + f) == False:
       break

2.图像预处理切片的问题

可以看到1中函数“ preprocess_mssd”中注释的部分，作为为通道减均值，注释掉了，下边还有直接整个图像减均值，两个操作区别导致模型图例卡了几天才找到问题。若执行通道减均值最后网络网络推理能得到不正常结果，若执行真个图像减去均值能得到正常结果。主要问题是读入图像时候图像数值类型+python切片造成的问题。 见下面demo：

import cv2
import numpy as np
img = cv2.imread("0_Parade_marchingband_1_156.jpg")
image = img.copy()
image = image[0:2,0:2,:]
image1 = image.copy()
image2 = image.copy()
print("origin image:",image.dtype)
print("origin image:\n",image)
image[:,:,0] = image[:,:,0] - 127.5
image[:,:,1] = image[:,:,1] - 127.5
image[:,:,2] = image[:,:,2] - 127.5
print("channel image: ",image.dtype)
print("channel image:\n",image)

image1 = image1 -127.5
print("mean image: ",image1.dtype)
print("mean image:\n",image1)

x = image2[:,:,0] - 127.5
image2[:,:,0] = image2[:,:,0] - 127.5
print("x image: ",x.dtype)
print("x image:\n",x)
print("image2 image: ",image2.dtype)
print("image2 image:\n",image2)

输出结果：

origin image: uint8
origin image:
 [[[67 68 78]
  [75 75 81]]

 [[71 72 82]
  [76 77 81]]]
channel image:  uint8
channel image:
 [[[196 197 207]
  [204 204 210]]

 [[200 201 211]
  [205 206 210]]]
mean image:  float64
mean image:
 [[[-60.5 -59.5 -49.5]
  [-52.5 -52.5 -46.5]]

 [[-56.5 -55.5 -45.5]
  [-51.5 -50.5 -46.5]]]
x image:  float64
x image:
 [[-60.5 -52.5]
 [-56.5 -51.5]]
image2 image:  uint8
image2 image:
 [[[196  68  78]
  [204  75  81]]

 [[200  72  82]
  [205  77  81]]]

可以看到，opencv读入图像为uint8类型，以切片方式赋值后类型不变origin image: uint8；切片方式减去float64类型数据后，在进行切片赋值，结果类型还是channel image: uint8，这就导致了间距之后越界，数据异常；整个图像间均值后赋值mean image: float64不会有这样的问题；后面的两个输x image: float64 和 image2 image: uint8出进一步说明了，切片后减均值会改变临时变量的数据类型为float64，但是再进行切片赋值后不会改边之前numpy数组的数据类型仍为int8，这就会导致减均值时候造成数据越界，得到异常结果。模型推理出问题整整卡了3天时间。