python - 如何解释我的模型是否发生过拟合或欠拟合以及如何提高我的准确性？

Question

我正在开发一种车牌识别算法。我的数据集包含 150 000 个训练、30.000 个测试和 680 个验证图像。我的模型结构是：

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
the_input (InputLayer)          (None, 64, 128, 3)   0                                            
__________________________________________________________________________________________________
conv1 (Conv2D)                  (None, 64, 128, 48)  3648        the_input[0][0]                  
__________________________________________________________________________________________________
pooling1 (MaxPooling2D)         (None, 32, 64, 48)   0           conv1[0][0]                      
__________________________________________________________________________________________________
dropout_1 (Dropout)             (None, 32, 64, 48)   0           pooling1[0][0]                   
__________________________________________________________________________________________________
conv2 (Conv2D)                  (None, 32, 64, 64)   76864       dropout_1[0][0]                  
__________________________________________________________________________________________________
pooling2 (MaxPooling2D)         (None, 16, 64, 64)   0           conv2[0][0]                      
__________________________________________________________________________________________________
dropout_2 (Dropout)             (None, 16, 64, 64)   0           pooling2[0][0]                   
__________________________________________________________________________________________________
conv3 (Conv2D)                  (None, 16, 64, 128)  204928      dropout_2[0][0]                  
__________________________________________________________________________________________________
pooling3 (MaxPooling2D)         (None, 8, 32, 128)   0           conv3[0][0]                      
__________________________________________________________________________________________________
dropout_3 (Dropout)             (None, 8, 32, 128)   0           pooling3[0][0]                   
__________________________________________________________________________________________________
flatten_1 (Flatten)             (None, 32768)        0           dropout_3[0][0]                  
__________________________________________________________________________________________________
dense_1 (Dense)                 (None, 160)          5243040     flatten_1[0][0]                  
__________________________________________________________________________________________________
dense_2 (Dense)                 (None, 64)           10304       dense_1[0][0]                    
__________________________________________________________________________________________________
out1 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out2 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out3 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out4 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out5 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out6 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out7 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out8 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
__________________________________________________________________________________________________
out9 (Dense)                    (None, 23)           1495        dense_2[0][0]                    
==================================================================================================
Total params: 5,552,239
Trainable params: 5,552,239
Non-trainable params: 0
________________________________________________________________________________ 
__________________

当我尝试从验证图像中识别车牌号时，它只能正确识别 18/680。我如何理解为什么训练和测试的准确率值很高，但是这个模型在训练后也不能识别验证图像和测试图像？如何改进我的模型？

它绘制了这样的图：

标签： pythontensorflowkerasdeep-learningcomputer-vision