Deep learning image recognition: Apple banana classification, are 458 pictures enough?

Deep learning image recognition: classification of apples and bananas, are 458 pictures enough?

This paper analyzes the sample size requirement for the recognition of apple and banana images using deep learning. In one case, using the ResNet50 model, 195 banana pictures and 263 apple pictures (458 in total) were collected respectively. After training, all pictures were misclassified as bananas, which raised questions about whether the sample size was sufficient.

The pre-trained ResNet50 model was used in the case, and the last fully connected layer was adjusted for binary classification. The code includes data augmentation techniques (such as random cropping and horizontal flip) and uses an SGD optimizer. However, due to the limited training samples and insufficient generalization capabilities of the model, the prediction results are severely wrong.

To this problem, an alternative is to extract image features using the pre-trained VGG16 model, and then use these features to train a three-layer multi-layer perceptron (MLP) for classification. This method believes that using the powerful feature extraction ability of pre-trained models can reduce the requirements for the number of training samples, and hundreds of pictures may be enough.

Therefore, in the case of limited data volume, it is crucial to choose the appropriate model architecture and feature extraction method. Large convolutional neural networks (such as ResNet50) are prone to overfitting training on small datasets, resulting in poor generalization ability. Using pre-trained models to extract features and then train, can effectively alleviate this problem and improve model performance.

The above is the detailed content of Deep learning image recognition: Apple banana classification, are 458 pictures enough?. For more information, please follow other related articles on the PHP Chinese website!