Research On Image Feature Recognition Method Based On Deep Learning

Image resources are becoming more accessible and more resourceful.In the modern manufacturing and smart life fields,using images is becoming increasingly common and important.But conventional recognizing methods only use single feature,and can't reflect the complex relationship between pixels sufficiently.Thus,they can not meet the increasingly diverse needs of practical application.Differently,deep learning method has the characteristics of automatic learning,excavating image information deeply and utilizing multi useful information comprehensively.It can improve the validity of feature recognition significantly when applied this method to the domain of image feature recognition.Three aspects on account of deep learning network are the focus of this study: the mechanism of deep image information excavation,the factors affecting its feature recognition performance,and the reasonable method of applying deep neural network to solve practical problems.Based on the fundamental theory of deep learning,the sparse auto-encoder deep network is designed in this paper.This network can effectively learn the elementary features used for image reconstruction.Through sparse combination of these elementary features,the useful information of the original image can be abstracted redundantly.The abstraction results can improve the effectiveness of image feature recognition better.Through experiments on handwritten digital images,the characteristics of sparse reconstruction of this deep network have been validated;meanwhile,through the control variate method,the related factors are studied,which influence the effect of deep network's image feature recognition.Experiments show that layer amounts and neuron amounts of each layer has the bell shaped curve characteristic.That is to say,setting too many or too few units can reduce the accuracy of image feature recognition.In additon,by using single maize seed images as subjects,the deep learning methods are studied for the kernel integrity feature recognition.Experiments are realized through sparse auto-encoder deep network and deep convolutionl neural network.The recognition accuracy of them has achieved up to 95%,and it is significantly higher than 71.93% attained by the traditional single hidden layer backpropagation neural network.When analysing the training time of these two networks comparatively,it can be seen that the learning speed of sparse auto-encoder is much faster,and this network can meet the needs of real-time more easily.Series of experiments prove that deep learning network is more suitable for solving the feature recognition problem of large scale images.And the SAE deep learning network adopted in this paper has the advantages of better recognition and faster training speed.When applied to the field of maize seed integrity feature recognition,this network can take into account both the accuracy and speed requirements.