Random identification error in the recognition program

The recognition program is always faced with a complex and diverse situation, in the process of improving the recognition performance, will encounter all kinds of mistakes, sometimes give people a kind of "impossible" feeling, on this, there is no two ways, only to strengthen the test data collection and reinforcement testing, but when we face a lot of mistakes, we will find that some of the random error That this error often occurs only once, it is difficult to reproduce, for this type of error, if the special method (that is, the solution is not the root of the problem) to solve it, the algorithm will not be adaptive, so there will be such a situation: although many recognition errors, but the final test after the recognition rate has not been improved. For random errors, the following four points are a more feasible strategy:

1, priority to solve the error of large probability, not entangled in small probability of the randomness of strong errors;

2, as far as possible from the randomness of errors found in the universality of the error, extrapolate to solve the problem;

3, properly tolerate the emergence of random errors, so as not to fall into the mire of algorithms;

4, strengthen the test, from the discrete error classification, summed up the error to find the law of error.

（这段时间在做识别算法性能提高，在这方面感悟较深，特拿出来与大家分享。）