Machine learning tasks such as image classification need to select the features that can describe the image well. The image has individual features and common features, and they are interdependent. If only the individual features of the image are emphasized, the neural network is prone to overfitting. If only the common features of images are emphasized, neural networks will not be able to adapt to diversified learning environments. In order to better integrate individual features and common features, based on skeleton and edge individual features extraction, this paper designed a mixed feature extraction method based on resonance filtering, named resonance layer. Resonance layer is in front of the neural network input layer, using K3M algorithm to extract image skeleton, using the Canny algorithm to extract image border, using resonance filtering to reconstruct training image by filtering image noise, through the common features of the images in the training set and efficient expression of individual characteristics to improve the efficiency of feature extraction of neural network, so as to improve the accuracy of neural network prediction. Taking the fully connected neural network and LeNet-5 neural networks for example, the experiment on handwritten digits database shows that the proposed mixed feature extraction method can improve the accuracy of training while filtering out part of image noise data.
Read full abstract