只保留图片重叠的区域来用来训练神经网络，网络的分辨准确率是应该上升还是下降？
比如训练一个分类mnist0和5的二分类网络
| 0 |
5 |
||||||||||||||||||
| 0 |
0 |
0 |
0.01 |
0.02 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0.11 |
0 |
0 |
||
| 0 |
0 |
0 |
0 |
0 |
0.03 |
0.09 |
0.07 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0.13 |
0 |
||
| 0 |
0 |
0 |
0.86 |
0.55 |
0.36 |
1 |
0.04 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0.33 |
0.59 |
0 |
||
| 0.02 |
0.01 |
0 |
0.72 |
0.05 |
0 |
0.12 |
0.94 |
0 |
0.05 |
0.02 |
0.07 |
0.05 |
0.83 |
0.89 |
0.01 |
1 |
0 |
||
| 0 |
0.05 |
0.73 |
0 |
0.04 |
0.02 |
0 |
1 |
0 |
0.02 |
0 |
0.05 |
0.96 |
0.94 |
0.87 |
0.04 |
1 |
0 |
||
| 0 |
0.01 |
0.76 |
0.96 |
0 |
0.05 |
0.05 |
0.83 |
0 |
0.01 |
0.06 |
0.95 |
0.66 |
0.05 |
0.95 |
0.1 |
1 |
0 |
||
| 0.04 |
0.04 |
0.03 |
1 |
0.91 |
0.06 |
0.09 |
0.93 |
0 |
0.02 |
0.04 |
1 |
0.11 |
0.1 |
0.66 |
1 |
0.68 |
0 |
||
| 0.07 |
0 |
0 |
0.91 |
0.39 |
0.15 |
0.89 |
0.93 |
0 |
0.02 |
0.05 |
0.67 |
0 |
0 |
0.02 |
0.07 |
0.01 |
0 |
||
| 0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
将图片从28*28处理成9*9，是每3个点取1个点。然后只保留两张图片重叠的区域
| 0 |
处理后 |
5 |
处理后 |
||||||||||||||||
| 0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
||
| 0 |
0 |
0 |
0 |
0 |
0 |
0 |
0.07 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0.13 |
0 |
||
| 0 |
0 |
0 |
0 |
0 |
0 |
1 |
0.04 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0.33 |
0.59 |
0 |
||
| 0.02 |
0.01 |
0 |
0.72 |
0.05 |
0 |
0.12 |
0.94 |
0 |
0.05 |
0.02 |
0 |
0.05 |
0.83 |
0 |
0.01 |
1 |
0 |
||
| 0 |
0 |
0.73 |
0 |
0.04 |
0.02 |
0 |
1 |
0 |
0 |
0 |
0.05 |
0 |
0.94 |
0.87 |
0.04 |
1 |
0 |
||
| 0 |
0.01 |
0.76 |
0.96 |
0 |
0.05 |
0.05 |
0.83 |
0 |
0 |
0.06 |
0.95 |
0.66 |
0 |
0.95 |
0.1 |
1 |
0 |
||
| 0.04 |
0.04 |
0.03 |
1 |
0.91 |
0.06 |
0.09 |
0.93 |
0 |
0.02 |
0.04 |
1 |
0.11 |
0.1 |
0.66 |
1 |
0.68 |
0 |
||
| 0.07 |
0 |
0 |
0 |
0 |
0.15 |
0.89 |
0.93 |
0 |
0.02 |
0 |
0 |
0 |
0 |
0.02 |
0.07 |
0.01 |
0 |
||
| 0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
||
处理后的效果如图。
剩余的4999组图片用同样的办法处理。用处理后的数据去训练网络，但测试集不变。
比较方法同样是用固定收敛标准多次测量取平均值的办法。
网络结构是
（mnist 0,mnist5）-81*30*2-(1,0)(0,1)
网络没有用卷积核，每个收敛标准计算199次。
比较正常输入训练的结果和经过重叠处理的结果
| 正常 |
重叠 |
正常 |
重叠 |
正常 |
重叠 |
正常 |
重叠 |
|||||
| δ |
迭代次数n |
迭代次数n |
平均准确率p-ave |
平均准确率p-ave |
耗时 min/199 |
耗时 min/199 |
最大值p-max |
最大值p-max |
||||
| 0.5 |
10.37688 |
8.954773869 |
0.862954 |
0.524822 |
0.513655 |
0.978722 |
0.05835 |
0.05905 |
1.011997 |
0.842415 |
0.747329 |
0.887127 |
| 0.4 |
260.804 |
500.9447236 |
1.920771 |
0.809037 |
0.802219 |
0.991572 |
0.0718 |
0.085817 |
1.195218 |
0.9375 |
0.956197 |
1.019943 |
| 0.3 |
342.5025 |
632.6633166 |
1.847179 |
0.903154 |
0.927125 |
1.026542 |
0.08035 |
0.096383 |
1.199544 |
0.94391 |
0.959936 |
1.016978 |
| 0.2 |
446.2211 |
731.8291457 |
1.640059 |
0.938246 |
0.950291 |
1.012837 |
0.08215 |
0.107383 |
1.307162 |
0.953526 |
0.966346 |
1.013445 |
| 0.1 |
545.2814 |
928.321608 |
1.702463 |
0.962159 |
0.955109 |
0.992674 |
0.0899 |
0.11475 |
1.276418 |
0.96688 |
0.963675 |
0.996685 |
| 0.01 |
858.794 |
1554.60804 |
1.810222 |
0.965648 |
0.963613 |
0.997893 |
0.116267 |
0.151117 |
1.299742 |
0.969551 |
0.96688 |
0.997245 |
| 0.001 |
1503.749 |
2944 |
1.957774 |
0.972279 |
0.972907 |
1.000646 |
0.1503 |
0.238017 |
1.583611 |
0.97703 |
0.974893 |
0.997813 |
| 9.00E-04 |
1570.111 |
2944 |
1.875027 |
0.972939 |
0.972695 |
0.999749 |
0.149817 |
0.232317 |
1.550673 |
0.978098 |
0.974359 |
0.996177 |
| 8.00E-04 |
1651.859 |
2944 |
1.782234 |
0.97324 |
0.972818 |
0.999567 |
0.158333 |
0.23075 |
1.457368 |
0.978098 |
0.974359 |
0.996177 |
| 7.00E-04 |
1721.397 |
2949.346734 |
1.713345 |
0.972977 |
0.972797 |
0.999815 |
0.16385 |
0.2321 |
1.41654 |
0.977564 |
0.974359 |
0.996721 |
| 6.00E-04 |
1791.035 |
2971.286432 |
1.658977 |
0.972252 |
0.972713 |
1.000475 |
0.168033 |
0.234433 |
1.39516 |
0.97703 |
0.974359 |
0.997266 |
| 5.00E-04 |
1852.794 |
3022.080402 |
1.631094 |
0.971575 |
0.972407 |
1.000856 |
0.1692 |
0.234 |
1.382979 |
0.97703 |
0.974893 |
0.997813 |
| 4.00E-04 |
2008.945 |
3306.170854 |
1.645725 |
0.971989 |
0.970907 |
0.998887 |
0.184333 |
0.2555 |
1.386076 |
0.978632 |
0.975427 |
0.996725 |
| 3.00E-04 |
2421.176 |
4384.321608 |
1.810823 |
0.974523 |
0.970818 |
0.996199 |
0.212933 |
0.341933 |
1.605823 |
0.980235 |
0.983974 |
1.003815 |
| 2.00E-04 |
2884.171 |
10441.28643 |
3.620204 |
0.978678 |
0.975862 |
0.997123 |
0.237933 |
0.697983 |
2.933525 |
0.981838 |
0.986645 |
1.004897 |
| 1.00E-04 |
3701.739 |
18356.21106 |
4.958808 |
0.978772 |
0.984849 |
1.006209 |
0.28295 |
1.208383 |
4.27066 |
0.980769 |
0.986645 |
1.005991 |
| 9.00E-05 |
3804.09 |
18591.38191 |
4.887208 |
0.978769 |
0.984519 |
1.005875 |
0.2906 |
1.384117 |
4.762962 |
0.980769 |
0.987179 |
1.006536 |
| 8.00E-05 |
4102.754 |
19038.83417 |
4.640501 |
0.978514 |
0.984286 |
1.005898 |
0.309067 |
1.4335 |
4.638158 |
0.981303 |
0.987714 |
1.006532 |
| 7.00E-05 |
4245.648 |
19696.58291 |
4.63924 |
0.978772 |
0.984353 |
1.005702 |
0.3175 |
1.44435 |
4.549134 |
0.981303 |
0.987714 |
1.006532 |
| 6.00E-05 |
4405.829 |
20854.70854 |
4.733436 |
0.978224 |
0.98453 |
1.006446 |
-1.84253 |
1.528483 |
-0.82956 |
0.981838 |
0.987714 |
1.005985 |
| 5.00E-05 |
4542.523 |
22543.76382 |
4.962829 |
0.976547 |
0.984514 |
1.008159 |
0.346 |
1.621517 |
4.686464 |
0.981838 |
0.988248 |
1.006529 |
| 4.00E-05 |
4640.01 |
25131.43216 |
5.416245 |
0.974501 |
0.985435 |
1.011219 |
0.331 |
1.7829 |
5.386405 |
0.981303 |
0.988248 |
1.007077 |
| 3.00E-05 |
4650 |
29251.83417 |
6.290717 |
0.974077 |
0.985861 |
1.012098 |
0.331617 |
2.065017 |
6.22712 |
0.975962 |
0.988248 |
1.012589 |
| 2.00E-05 |
4692.864 |
35816.31156 |
7.632079 |
0.974281 |
0.985386 |
1.011398 |
0.333417 |
2.42425 |
7.270932 |
0.979167 |
0.989316 |
1.010366 |
| 1.00E-05 |
5420.995 |
50387.91457 |
9.294957 |
0.97594 |
0.98526 |
1.00955 |
0.377167 |
3.283667 |
8.706142 |
0.981303 |
0.990385 |
1.009254 |
| 5.745602 |
1.008255 |
4.966842 |
1.007739 |
|||||||||
对比1e-4>=δ>=1e-5的区间
平均准确率上升0.8%，代价是耗时是原来的约5倍。因为将不重叠的区域去掉后分类性能是上升的，因此这个实验证明图片中重叠的区域才是决定分类性能的关键，而不重叠的区域事实上只是干扰。
| 重叠 |
05 |
无核 |
||||||
| f2[0] |
f2[1] |
迭代次数n |
平均准确率p-ave |
δ |
耗时ms/次 |
耗时ms/199次 |
耗时 min/199 |
最大值p-max |
| 0.500595 |
0.500735 |
8.954774 |
0.513655 |
0.5 |
17.79899 |
3543 |
0.05905 |
0.747329 |
| 0.539992 |
0.460466 |
500.9447 |
0.802219 |
0.4 |
25.86935 |
5149 |
0.085817 |
0.956197 |
| 0.581046 |
0.419118 |
632.6633 |
0.927125 |
0.3 |
28.9799 |
5783 |
0.096383 |
0.959936 |
| 0.479588 |
0.521407 |
731.8291 |
0.950291 |
0.2 |
32.36683 |
6443 |
0.107383 |
0.966346 |
| 0.796647 |
0.203773 |
928.3216 |
0.955109 |
0.1 |
34.57286 |
6885 |
0.11475 |
0.963675 |
| 0.013561 |
0.986429 |
1554.608 |
0.963613 |
0.01 |
45.54774 |
9067 |
0.151117 |
0.96688 |
| 4.68E-04 |
0.999533 |
2944 |
0.972907 |
0.001 |
71.74372 |
14281 |
0.238017 |
0.974893 |
| 4.65E-04 |
0.999536 |
2944 |
0.972695 |
9.00E-04 |
70.01005 |
13939 |
0.232317 |
0.974359 |
| 4.57E-04 |
0.999544 |
2944 |
0.972818 |
8.00E-04 |
69.50251 |
13845 |
0.23075 |
0.974359 |
| 4.64E-04 |
0.999537 |
2949.347 |
0.972797 |
7.00E-04 |
69.92462 |
13926 |
0.2321 |
0.974359 |
| 4.52E-04 |
0.999549 |
2971.286 |
0.972713 |
6.00E-04 |
70.63819 |
14066 |
0.234433 |
0.974359 |
| 4.11E-04 |
0.999589 |
3022.08 |
0.972407 |
5.00E-04 |
70.50754 |
14040 |
0.234 |
0.974893 |
| 3.07E-04 |
0.999694 |
3306.171 |
0.970907 |
4.00E-04 |
77.00503 |
15330 |
0.2555 |
0.975427 |
| 0.020341 |
0.979659 |
4384.322 |
0.970818 |
3.00E-04 |
103.0704 |
20516 |
0.341933 |
0.983974 |
| 0.241295 |
0.758705 |
10441.29 |
0.975862 |
2.00E-04 |
210.4322 |
41879 |
0.697983 |
0.986645 |
| 0.20105 |
0.79895 |
18356.21 |
0.984849 |
1.00E-04 |
364.2563 |
72503 |
1.208383 |
0.986645 |
| 0.226167 |
0.773833 |
18591.38 |
0.984519 |
9.00E-05 |
417.3116 |
83047 |
1.384117 |
0.987179 |
| 0.271384 |
0.728616 |
19038.83 |
0.984286 |
8.00E-05 |
432.206 |
86010 |
1.4335 |
0.987714 |
| 0.150792 |
0.849208 |
19696.58 |
0.984353 |
7.00E-05 |
435.4422 |
86661 |
1.44435 |
0.987714 |
| 0.15581 |
0.84419 |
20854.71 |
0.98453 |
6.00E-05 |
460.8291 |
91709 |
1.528483 |
0.987714 |
| 0.296499 |
0.703501 |
22543.76 |
0.984514 |
5.00E-05 |
488.8995 |
97291 |
1.621517 |
0.988248 |
| 0.527639 |
0.472361 |
25131.43 |
0.985435 |
4.00E-05 |
537.5578 |
106974 |
1.7829 |
0.988248 |
| 0.728633 |
0.271367 |
29251.83 |
0.985861 |
3.00E-05 |
622.5377 |
123901 |
2.065017 |
0.988248 |
| 0.834161 |
0.165839 |
35816.31 |
0.985386 |
2.00E-05 |
730.8894 |
145455 |
2.42425 |
0.989316 |
| 0.899491 |
0.100509 |
50387.91 |
0.98526 |
1.00E-05 |
990.0101 |
197020 |
3.283667 |
0.990385 |
| 正常 |
05 |
无核 |
||||||
| f2[0] |
f2[1] |
迭代次数n |
平均准确率p-ave |
δ |
耗时ms/次 |
耗时ms/199次 |
耗时 min/199 |
最大值p-max |
| 0.500303 |
0.499717 |
10.37688 |
0.524822 |
0.5 |
17.58794 |
3501 |
0.05835 |
0.842415 |
| 0.579083 |
0.42149 |
260.804 |
0.809037 |
0.4 |
21.57286 |
4308 |
0.0718 |
0.9375 |
| 0.562925 |
0.437137 |
342.5025 |
0.903154 |
0.3 |
24.19095 |
4821 |
0.08035 |
0.94391 |
| 0.716773 |
0.283868 |
446.2211 |
0.938246 |
0.2 |
24.76884 |
4929 |
0.08215 |
0.953526 |
| 0.158165 |
0.842016 |
545.2814 |
0.962159 |
0.1 |
27.09548 |
5394 |
0.0899 |
0.96688 |
| 0.09662 |
0.903382 |
858.794 |
0.965648 |
0.01 |
34.9196 |
6976 |
0.116267 |
0.969551 |
| 8.28E-04 |
0.99917 |
1503.749 |
0.972279 |
0.001 |
45.31658 |
9018 |
0.1503 |
0.97703 |
| 7.50E-04 |
0.99925 |
1570.111 |
0.972939 |
9.00E-04 |
45 |
8989 |
0.149817 |
0.978098 |
| 6.33E-04 |
0.999366 |
1651.859 |
0.97324 |
8.00E-04 |
47.73367 |
9500 |
0.158333 |
0.978098 |
| 5.23E-04 |
0.999478 |
1721.397 |
0.972977 |
7.00E-04 |
49.38191 |
9831 |
0.16385 |
0.977564 |
| 4.21E-04 |
0.999578 |
1791.035 |
0.972252 |
6.00E-04 |
50.66332 |
10082 |
0.168033 |
0.97703 |
| 3.52E-04 |
0.999648 |
1852.794 |
0.971575 |
5.00E-04 |
50.92462 |
10152 |
0.1692 |
0.97703 |
| 3.23E-04 |
0.999677 |
2008.945 |
0.971989 |
4.00E-04 |
55.47739 |
11060 |
0.184333 |
0.978632 |
| 2.59E-04 |
0.99974 |
2421.176 |
0.974523 |
3.00E-04 |
64.18593 |
12776 |
0.212933 |
0.980235 |
| 1.58E-04 |
0.999841 |
2884.171 |
0.978678 |
2.00E-04 |
71.72362 |
14276 |
0.237933 |
0.981838 |
| 8.69E-05 |
0.999913 |
3701.739 |
0.978772 |
1.00E-04 |
85.30653 |
16977 |
0.28295 |
0.980769 |
| 7.88E-05 |
0.999921 |
3804.09 |
0.978769 |
9.00E-05 |
87.58794 |
17436 |
0.2906 |
0.980769 |
| 6.85E-05 |
0.999932 |
4102.754 |
0.978514 |
8.00E-05 |
93.0201 |
18544 |
0.309067 |
0.981303 |
| 5.70E-05 |
0.999943 |
4245.648 |
0.978772 |
7.00E-05 |
95.71859 |
19050 |
0.3175 |
0.981303 |
| 4.35E-05 |
0.999957 |
4405.829 |
0.978224 |
6.00E-05 |
-555.543 |
-110552 |
-1.84253 |
0.981838 |
| 2.58E-05 |
0.999974 |
4542.523 |
0.976547 |
5.00E-05 |
104.3116 |
20760 |
0.346 |
0.981838 |
| 1.27E-05 |
0.999987 |
4640.01 |
0.974501 |
4.00E-05 |
99.76884 |
19860 |
0.331 |
0.981303 |
| 1.05E-05 |
0.999989 |
4650 |
0.974077 |
3.00E-05 |
99.9397 |
19897 |
0.331617 |
0.975962 |
| 1.07E-05 |
0.999989 |
4692.864 |
0.974281 |
2.00E-05 |
100.4171 |
20005 |
0.333417 |
0.979167 |
| 7.21E-06 |
0.999993 |
5420.995 |
0.97594 |
1.00E-05 |
113.6985 |
22630 |
0.377167 |
0.981303 |
