A deep learning method based on backpropagation

Abstract

The invention discloses a deep learning method based on backpropagation, comprising the following steps: S1: preparing a training set; S2: inputting the training set into a convolutional neural network to obtain a network output; S3: calculating the network output and the distance between the true values ​​in the training set to obtain the cross-entropy objective function; S4: judge whether the accuracy rate of the network output in step S2 is improved according to the cross-entropy objective function, if yes, then perform step S5, If not, then end the training; S5: use the sinusoidal exponential learning rate to update the weights of the convolutional neural network, and input the updated weights into the convolutional neural network in step S2, and then Steps S2 to S4 are repeated. The deep learning method proposed by the invention can greatly speed up the training of the classification network.

Description

technical field [0001] The invention relates to the fields of computer vision and image processing, in particular to a deep learning method based on backpropagation. Background technique [0002] Image classification belongs to the field of computer vision and image processing. It is one of the core problems in the field of computer vision. It has various practical applications. Many problems such as face positioning and pedestrian positioning can be attributed to classification problems, so image classification is a problem with Basic image processing problems have important academic and industrial research value. The goal of image super-resolution is to have a fixed set of classification labels, and then for the input image, find a classification label from the collection of classification labels, and finally assign the classification label to the input image. For humans, it is extremely simple to recognize a visual concept like a "cat". However, from the perspective of c...

AI Technical Summary

Problems solved by technology

[0003] The current mainstream image classification methods are mainly divided into three categories: methods based on the nearest neighbor classifier KNN, methods based on support vector machines, and methods based on deep learning; among them, the method based on KNN has simple and clear ideas and does not require training at all. However, it will be very time-consuming when the model is deployed, and it needs to be compared with all the pictures in the training set; the computational complexity of the method based on the support vector machine depends on the number of support vectors, not the dimension of the sample space. In this sense, the "curse of dimensionality" is avoided; however, its interpretation power for the high-dimensional mapping of the kernel function is not strong, and it depends on the experience of the experimenter, and is sensitive to data loss; the method based on deep learning learns a large number of priors from the training set. Knowledge can automatically extract features from data, so as not to rely on manual feature selection, so deep neural networks have achieved breakthrough results in computer vision, natural language processing, speech recognition and other fields; but in the process of use, The neural network hyperparameters involved require a long time to debug, and the training speed is generally slow, which affects the wide application of deep learning in image classification methods to a certain extent.

Images