Portable substance analysis based on computer vision, spectroscopy, and artificial intelligence
A portable complete analysis solution that integrates computer vision, spectrometry, and artificial intelligence for providing self-adaptive, real time information and recommendations for objects of interest. The solution has three major key components: (1) a camera enabled mobile device to capture an image of the object, followed by fast computer vision analysis for features and key elements extraction; (2) a portable wireless spectrometer to obtain spectral information of the object at areas of interest, followed by transmission of the data (data from all built in sensors) to the mobile device and the cloud; and (3) a sophisticated cloud based artificial intelligence model to encode the features from images and chemical information from spectral analysis to decode the object of interest. The complete solution provides fast, accurate, and real time analyses that allows users to obtain clear information about objects of interest as well as personalized recommendations based on the information.
1. A method for analyzing an object for one or multiple characteristic properties using a camera enabled device, a multi-wavelength spectrometer, and artificial intelligence comprising the steps of:
(a) obtaining an image of the object using a camera enabled device;
(b) using computer vision to extract one or more features of the object;
(c) using computer vision to identify one or more regions of interest within one or more of the features of the object;
(d) scanning the one or more regions of interest of the object using a multi-wavelength spectrometer over a plurality of wavelengths;
(e) measuring a spectrophotometric value for each of the wavelengths of step (d) to generate a spectrum of each region of interest of the object; and
(f) using artificial intelligence to integrate the spectrum generated in step (e) with the features extracted in step (b) with the regions of interest identified in step (c) to generate an analysis of the one or multiple characteristic properties of the object,
wherein the computer vision to extract one or more features of the object comprises:
downsampling the image of the object by performing a convolution with a filter bank using an encoder network;
reducing image resolution; and
inducing a set of feature maps of the image with the global and local features.
2. The method according to claim 1 wherein said one or more features of the object is a global feature or a local feature.
3. The method according to claim 1 wherein said spectrophotometric value is a high dimensional local feature.
4. The method according to claim 1 wherein in step (a) the image obtained by the camera enabled device is captured using a photon detection array for detecting light having a wavelength from about 200 nm to about 1100 nm.
5. The method according to claim 4 wherein the photon detection array is selected from a photo diode array or an active pixel sensor.
6. The method according to claim 1 wherein the typical encoder network comprises 8-16 convolutional layers in a VGG16 network.
7. The method according to claim 6 wherein the typical encoder network comprises 13 convolutional layers in a VGG16 network.
8. The method according to claim 1 wherein the plurality of wavelengths of steps (d) and (e) are selected from about 400 nm to about 1×10 4 nm.
9. The method according to claim 8 wherein the plurality of wavelengths of steps (d) and (e) are selected from about 400 nm to about 700 nm.
10. The method according to claim 8 wherein the plurality of wavelengths of steps (d) and (e) are selected from about 700 nm to about 1×10 4 nm.
11. The method according to claim 1 wherein the object is selected from humans, animals, plants, and inanimate image objects.
12. The method according to claim 11 wherein the feature is selected from water content, carbohydrate content, protein content, or color.
13. The method according to claim 1 wherein the computer vision to identify one or more regions of interest comprises:
upsampling the encoded feature maps for pixel-wise segmentation;
segmenting a region of interest of the object for spectroscopic measurements;
assigning a degree of interest on the region of interest of the object;
highlighting the region of interest of the object for spectroscopic measurements;
obtaining a spectrum on the region of interest of the object; and
condensing the image representation by a fully connected feedforward network.
14. The method according to claim 1 wherein the integration of the spectrum and the regions of interest further comprises:
inputting both the global and local features with the spectrum of the region of interest by a fusion multimodal probabilistic model; and outputting a final prediction of the characteristics of the object.
15. The method according to claim 14 wherein the multimodal probabilistic model comprises a group of classic discriminative classifiers, said classic discriminative classifiers selected from logistic regression, support vector machine, softmax classifier, or random forests for predicting a single label.
16. The method according to claim 14 wherein the multimodal probabilistic model further comprises a group of models simulating the dynamics of the object of interest and means making the prediction by considering the long-term structure within sequences.
17. The method according to claim 1 further comprising a fine-tuning procedure to achieve personalized implementation, wherein the fine-tuning procedure comprises:
implementing a standardized model for each user;
applying a matrix factorization to a collected dataset from the user;
leveraging the bioinformation of the user;
initiating a personalized recommendation;
acquiring both explicit and implicit feedback from the user; and
applying an optimization model to improve the quality of the personalized recommendation to an end user.
18. The method according to claim 17 wherein the optimization model is selected from stochastic gradient descent (SGD), root mean square propagation (RMSprop), adaptive gradient method (Adagrad), Adam optimizer (Adam), or contextual multi-arm bandit model.
19. A system according to claim 1 for analyzing an object for one or multiple characteristic properties, comprising:
(a) a processor;
(b) a storage medium;
(c) a spectrophotometer;
(d) a camera enabled mobile device; and
(e) a cloud based artificial intelligence,
wherein the spectrophotometer is operational over a plurality of wavelengths selected from about 400 nm to about 1×10 4 nm.
20. The system according to claim 19 , wherein the camera enabled mobile device comprises a photon detection array for detecting light having a wavelength from about 200 nm to about 1100 nm.
21. A method for analyzing an object for one or multiple characteristic properties using a camera enabled device, a multi-wavelength spectrometer, and artificial intelligence comprising the steps of:
(a) obtaining an image of the object using a camera enabled device;
(b) using computer vision to extract one or more features of the object;
(c) using computer vision to identify one or more regions of interest within one or more of the features of the object;
(d) scanning the one or more regions of interest of the object using a multi-wavelength spectrometer over a plurality of wavelengths;
(e) measuring a spectrophotometric value for each of the wavelengths of step (d) to generate a spectrum of each region of interest of the object; and
(f) using artificial intelligence to integrate the spectrum generated in step (e) with the features extracted in step (b) with the regions of interest identified in step (c) to generate an analysis of the one or multiple characteristic properties of the object,
wherein the computer vision to identify one or more regions of interest comprises:
upsampling the encoded feature maps for pixel-wise segmentation;
segmenting a region of interest of the object for spectroscopic measurements;
assigning a degree of interest on the region of interest of the object;
highlighting the region of interest of the object for spectroscopic measurements;
obtaining a spectrum on the region of interest of the object; and
condensing the image representation by a fully connected feedforward network.