Research On Classification Model Based On Hyperspectral And Deep Learning

The hyperspectral image has the characteristic of "map integration",and its data format is three-dimensional data cube.The high spectral imaging technology combines spectral analysis and imaging technology,and based on the near infrared spectral analysis,the image information of the sample is provided at the same time.In view of such data characteristics,the data processing of hyperspectral images can be analyzed in the spatial dimension,and the spectral data can be studied under the specified pixel.In re cent years,the simultaneous fusion of spectral dimension and spatial dimension data has become a research trend.Such data characteristics make hyperspectral gradually gain research and application in the quality inspection of various agricultural product s,especially fruits.In today's market,the detection of fruit quality is mainly carried out by traditional chemical methods,which must destroy the samples and consume time and energy.As for the data obtained through hyperspectral technology,the image information can be used to detect the external quality of fruits,such as whether there is damage,pollution,lesions,etc.,while the spectral information can be used to detect their internal quality and safety.In this paper,we set up the hardware system of hyperspectral on-line detection,and developed the software platform of hyperspectral acquisition system.Combined with the specific experimental requirements,the control of the spectral camera,high precision transmission belt and other key components in the online hyperspectral detection system was completed.At the same time,it can display the hyperspectral image in real time,save the detection data automatically,and display the data visually.Aiming at the problem of incomplete extraction of empty spectrum features of hyperspectral images,two kinds of deep learning models based on the information fusion of empty spectrum were proposed with strawberry as the experimental sample.In the first model,a two-branch network structure is adopted and a one-dimensional convolutional neural network is used to extract spectral dimension features of hyperspectral data.To extract the other branch of spatial morphology information,the data is firstly dimensionally reduced and then fed into the two-dimensional convolutional neural network,and the two branches finally carry out information fusion.On this basis,a classification model based on three-dimensional convolutional neural network is proposed to extract the characteristics of the combined space spectrum in hyperspectral images in order to extract the deep features of the deep fusion of space spectrum information.In this classification model,the 3d hyperspectral data containing the empty spectral characteristics of the sample can be directly input into the 3d convolutional neural network,which greatly preserves the original spatial structure of the target and avoids complex data reconstruction.Combined with the data characteristics of the specific sample strawberry,on this basis,the structure of the 3D convolutional neural network is further optimized.By connecting convolution layers with different convolution kernel sizes in the spectral dimension,a multi-scale 3D convolution module is formed.Based on this multi-scale 3D convolution module,a deep multi-scale 3D convolution neural network is formed to further improve the classification accuracy.At the same time,the sensitivity analysis of key hyperparameters is carried out for the two proposed models.