Light-weight regression network building method based on prior filtering

Abstract

The invention discloses a light-weight regression network building method based on prior filtering. The method comprises the steps of firstly performing a specified degeneration operation on each original image in an original image set, acquiring a corresponding degeneration image, slicing the original images and the corresponding degeneration images into image blocks, and acquiring training sample pairs; clustering in the training sample pairs, and dividing the training samples into different categories according to clustering results; for each category of sample pair, computing a prior filter of such category of samples, and performing three-value quantification on the prior filter; using the prior filters subjected to three-value quantification to build a light-weight regression networkand training, wherein after the light-weight regression network is trained, the image with high quality can be rebuilt in an end-to-end manner by using the input degenerated image. According to the method provided by the invention, the network with strong pertinence and lighter weight can be built by using fewer fixed prior filters, and the method is simple to compute, fast in training speed, small in needed storage space, and more applicable to a small sample problem.

Description

technical field [0001] The invention relates to the field of deep learning, in particular to a method for constructing a lightweight network based on a priori filter. Background technique [0002] Deep learning is the most promising direction in the field of machine learning. It is a hierarchical machine learning method including multi-level nonlinear transformation. By constructing a network structure suitable for different tasks, it has achieved far better results than traditional algorithms. In computer vision Breakthroughs have been made in areas such as object detection, natural language processing, speech recognition, and speech analysis. With the deepening of research, while the depth and size of the deep network model are approaching the accuracy limit of computer vision tasks, its depth and size are also increasing exponentially, and its required computing consumption and hardware costs are also increasing, which limits the depth. The network is widely used on mobi...

AI Technical Summary

Problems solved by technology

Recently, a fixed-parameter filter network Local Binary Convolution Neural Networks (LBCnn) based on LBP ideas uses LBP-based filters, but this network uses a large number of stacked filters to improve recognition accuracy, and the parameters are still massive.

Images