I was trying many tutorials for cnn that I've found online but they are not showing good performance. But my dataset is good enough, when I try with knn or mlp in sklearn I'm getting above 98% accuracy.
My images has 3 classes and has dimension of 100 by 100 px.
Please see the link:
https://www.learnopencv.com/image-classification-using-convolutional-neural-networks-in-keras/
https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html
Try Transfer Learning on Convolutional Neural Networks if you want better results. Train on VGG 16 (or VGG 19) Network with imagenet weights initially. First Freeze the hidden layers and train then after certain epochs, Unfreeze them and fine tune with either 2 Fully Connected layers or even logistic regression will work. The initial VGG16 (or 19) network will be your robust feature extractor.
for Conceptual understanding refer this : https://www.linkedin.com/pulse/transfer-learning-fine-tuning-lets-discuss-arun-das
More info: http://cs231n.github.io/transfer-learning/
Even more better insight : http://ruder.io/transfer-learning/
Reference Code : https://www.analyticsvidhya.com/blog/2017/06/transfer-learning-the-art-of-fine-tuning-a-pre-trained-model/
Transfer learning is expected to give you the accuracy you want.
Extra Note: After doing Feature extraction using VGG network, You can even use Machine learning classifiers (Like SVM) for the flattened vectors.
Could you share the domain of your application? What are you trying to classify? Also, what results you get with a CNN?
CNN is a kind of deep learning model and achieved promising results in image classification tasks. In other words, CNN acts as a powerful image classifier.
This link explains the usage of CNN as classifier for hand written digits.
https://naveenideas.blogspot.com/2018/12/convolutional-neural-network-using-keras.html
Hello Ruhshan
U can use VGG16(having 13 convolution layers and 3 fully connected layers) or vgg19 for classification of RGB images having 100*100 dimension in keras.
But nowadays GooLenet ,Resnet CNN architecture also become very powerful to train the large image data set for getting better accuracy or to minimize the error rate.
You can search here:
https://keras.io/examples/cifar10_resnet/
- Badges
- Science topic