Quick Notes

Experiment 1

Metric Learning，改变解码位置

Using Focal CTC Loss, 防止数据分布不均衡

ACE Loss, 顺序和频度问题, lambda=0.1

Experiment 2

分段式的CTC，避免双行车牌中首行字母的对齐问题

Experiment 3

flatten (reshape) 之后，使用 TCN

Accuracy: 67.00% Single Accuracy: 82.13% Double Accuracy: 7.41% Final Accuracy: 74.97%

在 Flatten 之前添加 LayerNorm

x = tf.split(f_map, num_or_size_splits=2, axis=1)
x = tf.concat(x, axis=2)
# layer norm
x = LayerNormalization()(x)
x = tf.reshape(x, (-1, 128, 168))

Accuracy: 70.25% Single Accuracy: 83.86% Double Accuracy: 16.67% Final Accuracy: 77.30%

直接对f_map进行reshape

x = tf.reshape(f_map, (-1, 128, 168))

Accuracy: 64.50% Single Accuracy: 73.51% Double Accuracy: 29.01% Final Accuracy: 82.69%

change to FasterNet

Accuracy: 61.38% Single Accuracy: 71.32% Double Accuracy: 22.22% Final Accuracy: 73.07%

Experiment 4

using segmentation framework, directly connect to the ctc

Accuracy: 60.75% Single Accuracy: 67.40% Double Accuracy: 34.57% Final Accuracy: 71.47%

Experiment 5

remove TCN model, only CNN, 先行预览版 for auto IT

减少不必要的结构, 改用 separable conv 连接 4x 的downsample

Accuracy: 86.12% Single Accuracy: 90.87% Double Accuracy: 67.88% Final Accuracy: 88.45%

Experiment 6

lraspp 连接 1/16 的 downsample 的 cnn

Accuracy: 87.62% Single Accuracy: 90.87% Double Accuracy: 75.15% Final Accuracy: 89.53%

Experiment 7

连接更多的浅层特征

Single Accuracy: 90.08% Double Accuracy: 64.24% Final Accuracy: 87.26%

Experiment 8

回归最淳朴的结构

Accuracy: 88.88% Single Accuracy: 91.18% Double Accuracy: 80.00% Final Accuracy: 90.34%

Experiment 9

repeat the train of double train and single double train

Original Accuracy: 90.38% S. LPR Accuracy: 91.50% D. LPR Accuracy: 86.06% Final Accuracy: 91.98%

Experiment 10

Focal CTC Loss(alpha=0.3, gamma=5.0), SGD optimizer

Original Accuracy: 91.75% S. LPR Accuracy: 91.34% D. LPR Accuracy: 93.33% Final Accuracy: 92.44%

Experiment 11

Focal CTC Loss(alpha=0.5, gamma=5.0), SGD optimizer

+-------------+-------+----------+ | | Count | Accuracy | +-------------+-------+----------+ | Single | 688 | 92.44% | | Double | 168 | 95.83% | | Total | 856 | 93.11% | | Error num. | 48 | 90.57% | | Error chars | 34 | 64.15% | | Final | 849 | 93.76% | +-------------+-------+----------+

on the test set

Original Accuracy: 95.66% S. LPR Accuracy: 95.71% D. LPR Accuracy: 95.39% Final Accuracy: 97.15%

noahzhy/LPR_v2