Image classification optimization method based on DARTS

Abstract

The invention discloses an image classification optimization method based on a DARTS, which is used for improving the stability and accuracy of a model constructed by the DARTS on image classification. According to the method, it is analyzed that jump connection operation has an unfair advantage in exclusive operation, and consequently DARTS performance collapse is caused, a new verification loss function is constructed in association with jump connection operation to hinder the unfair advantage, excessive jump connection is avoided, the architecture parameters are analyzed to determine a suitable sub-network for the particular image data set; and then an overall model training test is carried out, a certain number of searched unit architectures are stacked to construct a convolutional neural network, training is stared from the beginning, and a stable and accurate classification test is carried out on the image based on the model. According to the method, when specific images are classified, the unfair advantage of jump connection can be effectively hindered, and better performance can be generated.

Description

technical field [0001] The invention belongs to the field of deep learning and machine vision, and relates to a DARTS-based image classification optimization technology. This has a very wide range of applications in real-world scenarios, such as face recognition, fingerprint recognition, etc. Background technique [0002] Image classification is the task of extracting meaning from images using computer vision and machine learning algorithms. It seems simple, but this is one of the core issues in the field of computer vision, laying the technical foundation for problems in other vision fields (such as object detection and segmentation). Image classification tasks become difficult due to the diversity of image classification tasks. The total number of image categories is inconsistent, such as the dataset cifar10 and ImageNet’s 1000 categories; image features are diverse, such as mnist datasets with a single background and grayscale images and complex The background and the c...

AI Technical Summary

Problems solved by technology

In recent years, the two mainstream neural network architecture search methods with the best results are evolutionary algorithm (Evolutionary Algorithm) and reinforcement learning (Reinforcement learning), but these two architecture searches are equivalent to a black-box optimization problem in a discrete field, so resulting in the need to evaluate a large number of structures, very time-consuming

Images