End-to-end-for-chinese-plate-recognition
End-to-end-for-chinese-plate-recognition copied to clipboard
duanshengliu
cnn前做灰度化或二值化是否效果更好？
请教一个问题，我看代码cnn网络使用的是车牌图片修正后的3通道数据做输入，如果将输入图片做灰度化或二值化后再输入cnn是否训练速度更快，识别效果更好？
Model: "model"
Layer (type) Output Shape Param # Connected to
input_1 (InputLayer) [(None, 80, 240, 3)] 0
conv2d (Conv2D) (None, 80, 240, 16) 448 input_1[0][0]
max_pooling2d (MaxPooling2D) (None, 40, 120, 16) 0 conv2d[0][0]
conv2d_1 (Conv2D) (None, 38, 118, 32) 4640 max_pooling2d[0][0]
conv2d_2 (Conv2D) (None, 36, 116, 32) 9248 conv2d_1[0][0]
max_pooling2d_1 (MaxPooling2D) (None, 18, 58, 32) 0 conv2d_2[0][0]
dropout (Dropout) (None, 18, 58, 32) 0 max_pooling2d_1[0][0]
conv2d_3 (Conv2D) (None, 16, 56, 64) 18496 dropout[0][0]
conv2d_4 (Conv2D) (None, 14, 54, 64) 36928 conv2d_3[0][0]
max_pooling2d_2 (MaxPooling2D) (None, 7, 27, 64) 0 conv2d_4[0][0]
dropout_1 (Dropout) (None, 7, 27, 64) 0 max_pooling2d_2[0][0]
conv2d_5 (Conv2D) (None, 5, 25, 128) 73856 dropout_1[0][0]
conv2d_6 (Conv2D) (None, 3, 23, 128) 147584 conv2d_5[0][0]
max_pooling2d_3 (MaxPooling2D) (None, 2, 12, 128) 0 conv2d_6[0][0]
dropout_2 (Dropout) (None, 2, 12, 128) 0 max_pooling2d_3[0][0]
flatten (Flatten) (None, 3072) 0 dropout_2[0][0]
dropout_3 (Dropout) (None, 3072) 0 flatten[0][0]
c1 (Dense) (None, 65) 199745 dropout_3[0][0]
c2 (Dense) (None, 65) 199745 dropout_3[0][0]
c3 (Dense) (None, 65) 199745 dropout_3[0][0]
c4 (Dense) (None, 65) 199745 dropout_3[0][0]
c5 (Dense) (None, 65) 199745 dropout_3[0][0]
c6 (Dense) (None, 65) 199745 dropout_3[0][0]
c7 (Dense) (None, 65) 199745 dropout_3[0][0]
Total params: 1,689,415 Trainable params: 1,689,415 Non-trainable params: 0
二值化的阈值不太好设定，因为每张图片的彩色不一致，所以二值化以后图片车牌可能模糊化，效果不佳，会使得后续识别效果欠佳
------------------ 原始邮件 ------------------ 发件人: "duanshengliu/End-to-end-for-chinese-plate-recognition" <[email protected]>; 发送时间: 2020年11月8日(星期天) 中午11:44 收件人: "duanshengliu/End-to-end-for-chinese-plate-recognition"<[email protected]>; 抄送: "Subscribed"<[email protected]>; 主题: [duanshengliu/End-to-end-for-chinese-plate-recognition] cnn前做灰度化或二值化是否效果更好？ (#8)
请教一个问题，我看代码cnn网络使用的是车牌图片修正后的3通道数据做输入，如果将输入图片做灰度化或二值化后再输入cnn是否训练速度更快，识别效果更好？
Model: "model"
Layer (type) Output Shape Param # Connected to
input_1 (InputLayer) [(None, 80, 240, 3)] 0
conv2d (Conv2D) (None, 80, 240, 16) 448 input_1[0][0]
max_pooling2d (MaxPooling2D) (None, 40, 120, 16) 0 conv2d[0][0]
conv2d_1 (Conv2D) (None, 38, 118, 32) 4640 max_pooling2d[0][0]
conv2d_2 (Conv2D) (None, 36, 116, 32) 9248 conv2d_1[0][0]
max_pooling2d_1 (MaxPooling2D) (None, 18, 58, 32) 0 conv2d_2[0][0]
dropout (Dropout) (None, 18, 58, 32) 0 max_pooling2d_1[0][0]
conv2d_3 (Conv2D) (None, 16, 56, 64) 18496 dropout[0][0]
conv2d_4 (Conv2D) (None, 14, 54, 64) 36928 conv2d_3[0][0]
max_pooling2d_2 (MaxPooling2D) (None, 7, 27, 64) 0 conv2d_4[0][0]
dropout_1 (Dropout) (None, 7, 27, 64) 0 max_pooling2d_2[0][0]
conv2d_5 (Conv2D) (None, 5, 25, 128) 73856 dropout_1[0][0]
conv2d_6 (Conv2D) (None, 3, 23, 128) 147584 conv2d_5[0][0]
max_pooling2d_3 (MaxPooling2D) (None, 2, 12, 128) 0 conv2d_6[0][0]
dropout_2 (Dropout) (None, 2, 12, 128) 0 max_pooling2d_3[0][0]
flatten (Flatten) (None, 3072) 0 dropout_2[0][0]
dropout_3 (Dropout) (None, 3072) 0 flatten[0][0]
c1 (Dense) (None, 65) 199745 dropout_3[0][0]
c2 (Dense) (None, 65) 199745 dropout_3[0][0]
c3 (Dense) (None, 65) 199745 dropout_3[0][0]
c4 (Dense) (None, 65) 199745 dropout_3[0][0]
c5 (Dense) (None, 65) 199745 dropout_3[0][0]
c6 (Dense) (None, 65) 199745 dropout_3[0][0]
c7 (Dense) (None, 65) 199745 dropout_3[0][0]
Total params: 1,689,415 Trainable params: 1,689,415 Non-trainable params: 0
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or unsubscribe.
