基于多尺度特征和元学习的图像识别算法研究

IMAGE RECOGNITION ALGORITHMS WITH MULTISCALE FEATURES AND META-LEARNING

在计算机视觉与图像识别技术中,随着输入分辨率的变化,卷积神经网络方法的识别效果也不同。多尺度特征学习可以将图像的精度和细节结合在一起,结合图像的多个尺度信息进行分析。元学习让计算机模拟人的大脑,学习如何去学习,可以更高效灵活地实现图像分类。因此,结合多尺度特征和元学习进行图像识别算法研究具有较高的研究价值。通过膨胀卷积产生了不同分辨率的输入图像;使用元学习产生不同输入分辨率下神经网络的卷积权重;对于不同输入分辨率的模型使用知识蒸馏。

In computer vision and image recognition technology,the recognition effect of convolutional neural network methods varies with the change of input resolution.Multiscale feature learning can combine the accuracy and details of an image,combining the information of multiple scales of the image for analysis.Meta-learning allows computers to simulate the human brain and learn how to learn,which allows for more efficient and flexible image classification.Therefore,combining multi-scale features and meta-learning for image recognition algorithm research has high research value.In this study,we generated input images with different resolutions by inflated convolution,used meta-learning to generate convolutional weights of neural networks with different input resolutions,and used knowledge distillation for models with different input resolutions.