Image feature extraction algorithm based on parameter adaptive initialisation of CNN and LSTM

Image feature selection and extraction is the basis of image processing. For different types of images and different application requirements, the selection of image features will be different. The image feature extraction algorithm based on deep learning is a good method, but its overfitting, gradient explosion and initialisation of deep learning parameters are not adaptive, which directly leads to the low accuracy of the algorithm. Therefore, this paper proposes an image feature extraction algorithm based on adaptive initialisation of convolution neural network parameters based on MMN linear activation function. The parameter initialisation algorithm based on multi-layer Maxout activation function solves the problem of poor identification effect caused by improper parameter initialisation. Based on the selective Dropout algorithm of shallow learning of long and short time memory network, the problems of overfitting and gradient explosion in deep learning are overcome. SUSAN operator and random consistency algorithm are introduced to perform fine matching and purification. The adaptive parameter initialisation CNN algorithm proposed in this paper can effectively overcome the overfitting problem in the training process of deep learning, avoid the gradient explosion caused by improper parameter initialisation, and improve the calculation speed and accuracy of image feature extraction. And through the verification and simulation, the proposed image feature extraction algorithm can obtain more image details and edge information, and can better reflect the image feature information, for the image fusion and image recognition to lay a good foundation.