loss和val loss都在降低，但是分类模型的准度一直不能上升怎么办？

[BaTteREg]

310/310 - 68s - loss: 0.0710 - accuracy: 0.9958 - val_loss: 0.0652 - val_accuracy: 0.9940
Epoch 15/300
310/310 - 67s - loss: 0.0626 - accuracy: 0.9944 - val_loss: 0.0616 - val_accuracy: 0.9925
Epoch 16/300
310/310 - 67s - loss: 0.0505 - accuracy: 0.9968 - val_loss: 0.0524 - val_accuracy: 0.9950
Epoch 17/300
310/310 - 67s - loss: 0.0415 - accuracy: 0.9966 - val_loss: 0.0479 - val_accuracy: 0.9930
Epoch 18/300
310/310 - 67s - loss: 0.0517 - accuracy: 0.9927 - val_loss: 0.0538 - val_accuracy: 0.9915
Epoch 19/300
310/310 - 68s - loss: 0.0357 - accuracy: 0.9954 - val_loss: 0.0351 - val_accuracy: 0.9950
Epoch 20/300
310/310 - 67s - loss: 0.0239 - accuracy: 0.9978 - val_loss: 0.0351 - val_accuracy: 0.9920
Epoch 21/300
310/310 - 68s - loss: 0.0246 - accuracy: 0.9970 - val_loss: 0.0581 - val_accuracy: 0.9880
Epoch 22/300
310/310 - 68s - loss: 0.0318 - accuracy: 0.9937 - val_loss: 0.0338 - val_accuracy: 0.9925
Epoch 23/300
310/310 - 68s - loss: 0.0229 - accuracy: 0.9970 - val_loss: 0.0727 - val_accuracy: 0.9835
Epoch 24/300
310/310 - 69s - loss: 0.0237 - accuracy: 0.9966 - val_loss: 0.0349 - val_accuracy: 0.9920
Epoch 25/300

主要是验证集的准度一直不能提升到0.9950以上
训练环境几乎没有噪声影响，我看别的论文都能做到100%
求大佬帮助！