Research On Classification Of Freshwater Algae Images Based On Improved Convolution Neural Network

Algae plants,the main primary producers in nature,play an important role in the material circulation and energy flow of the whole natural ecosystem.With the increasing environmental damage,water pollution has become a major problem to be solved urgently.When the water is polluted,such as eutrophication,it will cause the rapid propagation of some algae,the bloom and the red tide.These phenomena will greatly destroy the balance of matter and energy in the water,and eventually destroy the whole water ecosystem.Therefore,the distribution of algae is an important indicator to judge the environmental pollution of water,and it is very important for the study of algae detection.At present,the detection methods of algae mainly rely on artificial recognition,low efficiency and great labor intensity.In the thesis,we use convolution neural network to classify and identify algal atlas,and improve the convolution neural network model according to the characteristics of algal diversity.The specific research work is as follows:1.The theory of digital lensless holographic imaging is introduced.Based on the coaxial imaging method,the algae detection device is made by using 3D printer.In view of the algal holographic diffraction images collected by CCD,the influence of the conjugate image iteration algorithm to eliminate the conjugate image is proposed.finally,16000 algae collections of freshwater algae were collected by using the algal device.2.A robust adaptive stochastic gradient optimization algorithm is proposed.A robust adaptive learning rate optimization algorithm is proposed for convolution neural network(CNN)learning process based on huge data computation.The random curvature rate information is used to automatically adjust the loss function of the learning rate.3.The principal component analysis pooling method(PCA-pooling)is proposed.For the convolution neural network,we also discard a large number of image feature information during the pooling process.In combination with the ?1 norm loss function,we use the augmented Lagrangian operator to calculate the optimal output weight by evaluating the structural machine learning(ELM)model structural risk and empirical risk.This results in a more robust model,faster calculation rate,and accurate classification of the ?1-ELM model.4.We present a CNN-?1-ELMclassification model.The algorithm uses CNN to perform feature extraction of input complex algae images on an adaptive stochastic gradient convolutional neural network,and combines ?1-ELM as a classifier with high generalization and fast learning rate.Finally,according to the actual collection of algae atlas,select appropriate parameters,train the model,and classify the algae and the final recognition rate of 93.6% is reached.