Abstract: Aiming at the problem that the  current convolution kernel initialization method was easy to lead to network instability and the limitation of principal component analysis algorithm on network structure, we proposed a convolution kernel initialization method based on image features. Firstly, the method combined fuzzy processing technology and edge processing technology to sample images, and then the sampled data were randomly divided into groups. Principal component analysis algorithm was used to extract the principal components of each group of data, and the convolution kernel was initialized. We applied the method to Cifar-10 and Corel-1000 datasets, and compared it with Gaussian initialization method and He initialization method. The experimental results show that the performance of the method is superior to other convolution kernel initialization methods.

Key words: deep learning, convolution kernel initialization, image feature, principal component analysis, random combination