|
Recognition methods | Main basis | Typical methods | Advantages | Disadvantages |
|
Recognition method based on template matching | Global property | Traversal search matching algorithm | The program is simple and suitable for standard character verification code. | The required template library is large; it depends on the choice of template matching. |
Local feature | Shape context matching algorithm | The image information is rich, and it is robust to image scaling and affine transformation. | Without rotation invariant. |
|
Recognition method based on character feature | Character structure feature | Algorithm based on character structure feature | Sensitive to the details of characters; strong in distinguishing features. | The distortion is serious when there are noise interferences. |
Character statistical feature | Algorithm based on character statistical feature | Strong robustness against noise interference. | Targeted; application limited. |
|
Recognition method based on machine learning | Template matching | SVM | Strong approximation ability and generalization ability; good adaptability and high accuracy for small sample space; suitable for two kinds of classification. | Not applied to infinite sample space. |
KNN | It is better to avoid the problem of imbalanced samples, which is suitable for overlapping samples of the same class. | Computation is complex; easy to misjudge in the domain with small sample size. |
Traditional method | BPNN | Flexible structure design, suitable for multiclass classification. | Slow convergence rate; depends on parameters. |
Deep learning | CNN | Accepts an input image directly; automatically extracts features; own robustness to displacement, scale, and deformation; high recognition accuracy. | Lack of time dimension; it could not identify using context information. |
RNN | Processes data in time series. | Time gradient may disappear. |
LSTM-RNN | Owns the time memory function; effective to prevent gradient disappear. | Unable to extract feature automatically. |
|