Image Representation Learning Based On Convolutional Auto-encoders

In the research of computer vision,how to obtain representative abstract image features is the key point to make computer help people to understand semantic mean-ings of images.Traditional image feature extracting methods are mostly based on the human-beings' understanding of images and hand-crafting features,hence they are quite limited.Nowadays,using deep representation learning methods represented by convolutional neural network to learn image representations automatically has grad-ually become the mainstream research direction.However,training deep neural net-works often needs a large amount of labeled data,which are quite expensive to obtain.Hence how to do representation learning based on neural network in semi-supervised or weakly supervised manner becomes a important problem to be solved.In this pa-per,we study convolutional neural network based denoising auto-encoders,explore its representation learning effect in self-supervised and weakly supervised learning,give solutions when hardware resources are limited,and discuss its application on micro-scopic image recognition problem.The main contributions of this paper are:Firstly,aiming at the problem that images have less standard data,a self-supervised learning method that can use non-classical data to learn is proposed.This method use convolutional auto-encoder network with symmetrical skip connections.It can learn image representations through its encoder and decoder architecture.With the added symmetrical skip connections,it can bypass detail information to the bottom layers of the network to learn more abstract representations when it learns to solve low-level image problems.Secondly,aiming at the efficiency of deep convolutional self-encoder,we pro-pose an auxiliary training method which can use deep auto-encoder network and deep classification network to help to train shallower auto-encoder network.Hence it can maximize the effect of restoring clean image and learning good representations while accelerate their training efficiency.Thirdly,we apply the convolutional auto-encoder network to fungal microscopic image recognition problem.For the case using all labeled images and partially labeled images,we both get more accuracy than 95%.