Research On Improved Algorithm Of SAE And Its Application In Rice Protein Content Detection Based On Hyperspectral Images

In recent years,with the development of AI,deep learning algorithm has attracted more and more attention of professional scholars.Auto-encoder(AE),a typical deep learning algorithm,have many corresponding improved algorithms.Among them,stacked auto-encoder(SAE)acquired by cascading multiple AEs has been widely used in dimensionality reduction of high-dimensional data,especially in the processing of non-linear data.However,it is impossible to evaluate whether the information extracted by SAE contains noise or not.Low rank matrix recovery(LRMR)algorithm obtains the low rank components of data by matrix decomposition,so as to realize data denoising.In this paper,low rank auto-encoder(LR-SAE),an improved algorithm of auto-encoder,combines the dimension reduction advantage of SAE and the noise reduction advantage of LRMR algorithm,which was proposed to improve the algorithm performance of auto-encoder.China is a big rice producer and consumer.Protein content is an important indicator of the nutritional value of rice.Hyperspectral image technology can realize the non-destructive detection of rice protein content,but the huge amount of hyperspectral image data and high redundancy restricts the subsequent analysis results.Although the traditional linear dimension reduction methods can extract the feature of hyperspectral image data well,they cannot meet the requirement of non-linear data dimension reduction,and the noise of hyperspectral image also restricts the final analysis results.In this paper,in order to solve the above problems,LR-SAE was used to reduce the dimension of hyperspectral image of rice samples,and support vector machine regression(SVR)was used to construct the analysis model,which achieved fast,accurate and non-destructive rice protein content detection,and verified the efficiency of LR-SAE.The main content and conclusions in this research are as following:(1)The principle of SAE was studied.AE has been widely used in the field of data dimension reduction,because of its excellent feature extraction ability.SAE greatly improves the performance of this algorithm by stacking and cascading multiple AEs.The application of LRMR algorithm in data denoising is becoming more and more mature.In this paper,LR-SAE algorithm was proposed by combining SAE and LRMR algorithm.Before the training of each hidden layer in the SAE network,a low rank decomposition layer is added to extract the low rank or approximate low rank structure of each layer of neuron data.The structure is considered to be low-noise,then the corresponding hidden layer data is trained.The network is trained by greedy layer-wise training method,and the neuron information of the last hidden layer,i.e.the deep robust feature of training data,is finally extracted.(2)Total 420 hyperspectral images(400-1000 nm)of rice samples were collected.The average spectral value in the region of interest(ROI)was extracted as the spectral information of the sample.Simultaneously,the images of the respective bands in ROI were separated,and a total of 478 ordinary images were obtained as the image information of the sample.(3)From the perspectives of spectral information,image information and spectral-image fusion information,feature extraction was performed by principal component analysis(PCA),SAE and LR-SAE.First,the prediction model was built on the spectral feature.Before feature extraction,SG preprocessing was applied to the original hyperspectral data to reduce the effects of various noises during the acquisition process.Then,based on the low-dimensional feature after feature extraction,the prediction model of rice protein content was established by using SVR.The results show that the model based on LR-SAE and SVR has the best prediction effect,_C~2 is 0.9926,RMSEC is 0.0437,_P~2 is 0.9394,RMSEP is 0.1232.Second,the prediction model was built based on the image feature.The RGB images was transformed into gray scale images of 28*28 pixels.The gray scale images were further flattened and transformed into 784-dimensional column vectors.Then,based on the low-dimensional feature after feature extraction,SVR was used to establish a prediction model of rice protein content.The results show that the model based on LR-SAE and SVR has the best prediction effect,_C~2 is 0.9569,RMSEC is0.0860,_P~2 is 0.8769,and RMSEP is 0.3394.Third,the prediction model was established based on the fusion feature of spectral-image fusion information.The 478-dimensional spectral information and 784-dimensional image information were organically fused to form 1262-dimensional spectral-image fusion information.Based on the low-dimensional feature after feature extraction,the prediction model of rice protein content was established by SVR.The results show that the model based on LR-SAE and SVR has the best prediction effect,_C~2 is 0.9931,RMSEC is 0.04,_P~2 is 0.9619,and RMSEP is 0.0854.Compared with PCA and SAE,feature extraction based on LR-SAE performed best,which verified the advantages of LR-SAE.The results of the modeling based on three different perspectives show that the effect of the model based on spectral-image fusion information is the best among the three models.The analysis model based on the fusion of spectral and image information of hyperspectral image can realize fast,accurate and non-destructive detection of rice protein content with higher efficiency.