2

我正在尝试矢量化滑动窗口搜索以进行对象检测。到目前为止，我已经能够使用 numpy 广播将我的主图像切成窗口大小的切片，这些切片存储在all_windows下面看到的变量中。我已经验证了实际值是否匹配，所以我对此感到满意。

下一部分是我遇到麻烦的地方。我想在all_windows调用函数时对数组进行索引，patchCleanNPredict()以便可以以类似的矢量化格式将每个窗口传递给函数。

我试图创建一个名为 new_indx 的数组，该数组将包含二维数组中的切片索引，例如 ([0,0], [1,0], [2,0]...)，但遇到了问题。

我希望最终得到每个窗口的一系列置信度值。下面的代码适用于 python 3.5。提前感谢您的任何帮助/建议。

import numpy as np

def patchCleanNPredict(patch):
    # patch = cv2.resize()# shrink patches with opencv resize function
    patch = np.resize(patch.flatten(),(1,np.shape(patch.flatten())[0])) # flatten the patch
    print('patch: ',patch.shape) 
    # confidence = predict(patch) # fake function showing prediction intent
    return # confidence


window = (30,46)# window dimensions
strideY = 10
strideX = 10

img = np.random.randint(0,245,(640,480)) # image that is being sliced by the windows

indx = np.arange(0,img.shape[0]-window[1],strideY)[:,None]+np.arange(window[1])
vertical_windows = img[indx]
print(vertical_windows.shape) # returns (60,46,480)


vertical_windows = np.transpose(vertical_windows,(0,2,1))
indx = np.arange(0,vertical_windows.shape[1]-window[0],strideX)[:,None]+np.arange(window[0])
all_windows = vertical_windows[0:vertical_windows.shape[0],indx]
all_windows = np.transpose(all_windows,(1,0,3,2))

print(all_windows.shape) # returns (45,60,46,30)


data_patch_size = (int(window[0]/2),int(window[1]/2)) # size the windows will be shrunk to

single_patch = all_windows[0,0,:,:]
patchCleanNPredict(single_patch) # prints the flattened patch size (1,1380)

new_indx = (1,1) # should this be an array of indices? 
patchCleanNPredict(all_windows[new_indx,:,:]) ## this is where I'm having trouble
4

1 回答 1

0

为了以矢量化方式评估所有窗口上的函数，我最终不得不使用 np.transpose 进行大量调整大小和重新排列以使其全部正确广播。下面的代码有效，并且有 for 循环来显示并确认图像窗口没有乱码/混淆。它们将被删除/评论以进行全速运行。

一个小的免责声明：我认为必须有更清洁的跨 2D 矩阵的滑动窗口实现，但由于我找不到任何下面的示例可能对其他人有所帮助。此外，如果对广播语法有更透彻的了解，可能会清理一些频繁的重新排列和调整大小。

import numpy as np
import cv2


def Predict(flattened_patches):
    # taking the mean of the flattened windows and then returning the
    # index of the row (window) with the highest mean, a predicter would have the same syntax
    results = flattened_patches.mean(1) 
    max_index = results.argmax() 
    return results, max_index

## -------- image and sliding window setup -------------------------
AR = 1.45 # choose an aspect ratio to maintain throughout scaling steps
win_h = 200 # window height
win_w = int(win_h/AR) # window width
window = (win_w,win_h)# window dimensions
strideY = 100
strideX = 100

data_patch_size = (30,46) # size the windows will be shrunk to for object detection

img = cv2.imread('picture6.png') # load an image to slide over

cv2.namedWindow('image',cv2.WINDOW_NORMAL) 
cv2.resizeWindow("image",int(img.shape[1]/2),int(img.shape[0]/2)) # shrink the image viewing window if you have large images

img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
## -------- end of, image and sliding window setup --------------------

## -------- sliding window vectorization steps --------------------------
num_vert_windows = len(np.arange(0,img.shape[0]-window[1],strideY)) # number of vertical windows that will be created
indx = np.arange(0,img.shape[0]-window[1],strideY)[:,None]+np.arange(window[1]) # index that will be broadcasted across image
vertical_windows = img[indx] # array of windows win_h tall and the full width of the image

vertical_windows = np.transpose(vertical_windows,(0,2,1)) # transpose to prep for broadcasting
num_horz_windows = len(np.arange(0,vertical_windows.shape[1]-window[0],strideX)) # number of horizontal windows that will be created
indx = np.arange(0,vertical_windows.shape[1]-window[0],strideX)[:,None]+np.arange(window[0]) # index for broadcasting across vertical windows
all_windows = vertical_windows[0:vertical_windows.shape[0],indx] # array of all the windows
## -------- end of, sliding window vectorization ------------------------

## ------- The below code rearranges and flattens the windows into a single matrix of pixels in columns and each window
## ------- in a row which makes evaluating a function over every window in a vectorized manner easier

total_windows = num_vert_windows*num_horz_windows

all_windows = np.transpose(all_windows,(3,2,1,0)) # rearrange for resizing and intuitive indexing

print('all_windows shape as stored in 2d matrix:', all_windows.shape)
for i in range(all_windows.shape[2]): # display windows for visual confirmation
    for j in range(all_windows.shape[3]):
        cv2.imshow('image',all_windows[:,:,i,j])
        cv2.waitKey(100)

all_windows = np.resize(all_windows,(win_h,win_w,total_windows))
print('all_windows shape after folding into 1d vector:', all_windows.shape)
for i in range(all_windows.shape[2]): # display windows for visual confirmation
    cv2.imshow('image',all_windows[:,:,i])
    cv2.waitKey(100)

# shrinking all the windows down to the size needed for object detect predictions
small_windows = cv2.resize(all_windows[:,:,0:all_windows.shape[2]],data_patch_size,0,0,cv2.INTER_AREA)
print('all_windows shape after shrinking to evaluation size:',small_windows.shape)
for i in range(small_windows.shape[2]): # display windows for vis. conf.
    cv2.imshow('image',small_windows[:,:,i])
    cv2.waitKey(100)

# flattening and rearranging the window data so that the pixels are in columns and each window is a row
flat_windows = np.resize(small_windows,(data_patch_size[0]*data_patch_size[1],total_windows))
flat_windows = np.transpose(flat_windows)
print('shape of the window data to send to the predicter:',np.shape(flat_windows))

results, max_index = Predict(flat_windows) # get predictions on all the windows
于 2016-11-24T04:08:03.437 回答