python numpy 显示图像阵列的实例
每次要显示图像阵列的时候,使用自带的 matplotlib 或者cv2 都要设置一大堆东西,subplot,fig等等,突然想起 可以利用numpy 的htstack() 和 vstack() 将图片对接起来组成一张新的图片。因此写了写了下面的函数。做了部分注释,一些比较绕的地方可以自行体会。
大致流程包括:
1、输入图像列表 img_list
2、show_type : 最终的显示方式,输入为行数列数 (例如 show_type=22 ,则最终显示图片为两行两列)
3、basic_shape, 图片resize的尺寸。
def image_show( img_list, show_type, basic_size=[300,500]): ''' img_list contains the images that need to be stitched, the show_typ contains the final shape of the stitched one, ie, 12 for 1 row 2 cols. basic_size : all input image need to be reshaped first. ''' # reshap row and col number. n_row, n_col = basic_size #print n_row,n_col # num of pixels need to be filled vertically and horizontally. h_filling = 10 v_filling = 10 # image resize. resize_list=[] for i in img_list: temp_img = cv2.resize( i, ( n_col, n_row ), interpolation = cv2. INTER_CUBIC ) resize_list.append( temp_img ) # resolve the final stitched image 's shape. n_row_img, n_col_img = show_type/10, show_type%10 #print n_row_img, n_col_img # the blank_img and the image need to be filled should be defined firstly. blank_img= np.ones([n_row,n_col])*255 blank_img= np.array( blank_img, np.uint8 ) v_img= np.array( np.ones([n_row,v_filling])*255, np.uint8) h_img= np.array( np.ones ([ h_filling, n_col_img*n_col+(n_col_img-1)*h_filling])*255, np.uint8) # images in the image list should be dispatched into different sub-list # in each sub list the images will be connected horizontally. recombination_list=[] temp_list=[] n_list= len(resize_list) for index, i in enumerate ( xrange (n_list)): if index!= 0 and index % n_col_img==0 : recombination_list.append(temp_list) temp_list = [] if len(resize_list)> n_col_img: pass else: recombination_list.append(resize_list) break temp_list.append( resize_list.pop(0)) if n_list== n_col_img: recombination_list.append(temp_list) #print len(temp_list) #print temp_list # stack the images horizontally. h_temp=[] for i in recombination_list: #print len(i) if len(i)==n_col_img: temp_new_i=[ [j,v_img] if index+1 != len(i) else j for index, j in enumerate (i) ] new_i=[ j for i in temp_new_i[:-1] for j in i ] new_i.append( temp_new_i[-1]) h_temp.append(np.hstack(new_i)) else: add_n= n_col_img - len(i) for k in range(add_n): i.append(blank_img) temp_new_i=[ [j,v_img] if index+1 != len(i) else j for index, j in enumerate (i) ] new_i=[ j for i in temp_new_i[:-1] for j in i ] new_i.append( temp_new_i[-1]) h_temp.append(np.hstack(new_i)) #print len(h_temp) #print h_temp temp_full_img= [ [j, h_img ] if index+1 != len(h_temp) else j for index, j in enumerate(h_temp) ] if len(temp_full_img) > 2: full_img= [ j for i in temp_full_img[:-1] for j in i ] full_img.append(temp_full_img[-1]) else: full_img= [ j for i in temp_full_img for j in i ] #full_img.append(temp_full_img[-1]) if len(full_img)>1: return np.vstack( full_img) else: return full_img
最终输入情况和结果如下图:
第一组结果图:自行看输入
第二组结果图。
以上这篇python numpy 显示图像阵列的实例就是小编分享给大家的全部内容了,希望能给大家一个参考,也希望大家多多支持【听图阁-专注于Python设计】。