DETAILED DESCRIPTION OF THE INVENTION
This invention is illustrated in the accompanying drawings, throughout which, numerals
indicate the corresponding parts in the various figures.
The PatcntNo.513925 dated Feb 22,2024 uses CNN for image recognition and traditional
microscope to capture images.
The current invention has two maJor parts. They arc i) uses MSFE light weight
architecture for image recognition and fold scope for image capturing i i) features like QR code
and voice assistant integration helps in enhanced user experience.
TECHNOLOGY BElliNI> THE PRODUCT:
Figure I shows the assembled foldscopc. Foldscope is an optical microscope that can be
assembled using a sheet of' paper and lens. This foldscope can be assembled in less than seven
minutes. Once assembled, it is about the size of the bookmark.
As scm bling Foldscopc:
The Foldscope is cheap and easy to usc tool for scientific purpose. It is an ultra-low-cost
origami-based approach for large scale manufacturing of microscopes. It is made of water-resistant
paper. The basic principle behind this f'oldscope is using a small spherical lens which is held close
to eyes. The foldscopc weighs about 8 grams that comes in a kit with a lens that magnifies 140X.
The Figure 2 shows what the kit contains. The kit also contains magnets i.e.: couplers that can be
used to allach the l'oldscopc with the smartphonc. The focusing ramp helps to move left/right to
locus by placing our fingers on Trace and Grip holes that looks like a half-moon. LED magnifier
can be used as a light source while viewing our micro cells in which it is placed on apenurc. The
slide can be fixed behind the sample stage .
Block Diagram:
Figure 3 shows the block diagram of the system. It consists of the five modules. They arc:
I) Image Capturing:
2) In this module, image is either captured from foldscope or it can be used from memory
unit. Here, we can also perform virtual staining process which is helpful for research
purposes.
3) Image Verification:
Here, captured image is verified and checked if it is compatible. If it is compatible it can
proceed to next module, else, we need to capture another image.
4) Image Recognition:
In this module, we need to classify the image with the help of deep learning techniques.
The classilication will be based on trained model.
5) Image Segmentation:
The recognized image is then segmented to calculate the starch percentage of the potato
tuber which is the most important and crucial.part of the system.
6) Quality Report Generation:
Outputs from 1·ccognition module and segmentation module is then f'cd into to this module.
This contains 2 important sections, one is QR code generation for the generated quality
report and another one is integrating voice assistant.
V~:RIFICATION OF IMAGE:
The verification algorithm works based on different mctrics. The four attributes of this
algorithm arc blurriness, focus quality, bright spot and texture of images. 131urrincss is calculated
\Ising l.arlacian variance. The GLCM (Gabor Filter and Gray Level Matrix) features like
Homogeneity, Dissimilarity, Energy and Entropy helps to identify the spacial texture of the
images. Mean of the Gabor filter helps to identify the spectral texture of the image. Spectral
Moments arc used to identify the focus quality. Also, bright spot of the image is detected using
Otsu's thresholding Method. _We find variance of the binary image by arrlying Otsu's
Thresholding. Then, bright spot is detected based on this variance. We strictly check for blurriness
using the Laplacian Variance. Atleast six features are needed to get compatible image. If not, we
need to select another image and check for compatibility. The compatible image will be fed into
Image Recognition module.
I'IWI'OSED IH:EI' U:ARNING MSFE ARCHITECTURE:
The image recognition has been performed using deep learning techniques. To make it
more resource constraints of the edge devices like smartphone, light-weight architecture along
with MSFE, a Multi Scale Feature Extraction has been designed. MSFE has DWS (Depth wise
Convolution) and residual connections. It helps in extracting complex features. It diminishes the
size and computational cost of the networks by improving the performance. Additionally, for the
smooth gradient flow, we have used residual connections which ensures stable training. The
proposed network consists of 4 MSFE blocks followed by Average Pooling and Flattening along
with residual connections. Aficr this, images will be fed into dense layer where sigmoid/softmax
will help to identify the core type and weight of the potato tuber. Image Recognition will be done
under the multi-class or multi-label classification. The multi-class classification will output the
size of the tuber (Large, Medium, Small) while multi-label classification will output core type of
the tuber (Inner Core or Outer Core). To improve network performance, a - balanced Focal Loss has been utilized and Binary Cross Entropy has been utilized to make the model pay more attention
to the minority class. Also, RMSI'rop has been adopted to improve the performance and speed of
the training models.
IMAGE SEGMENTATION:
Firstly, we will preprocess and binarize the image. Then the image will be segmented by
performing some morphological operations. We use clahe technique and apply black hat
transformation for preprocessing and then we concatenate those output images to enhance it. Using
Otsu's Thresholding, we binarizc the image. Using that binary image, we perform morphological
operations like eroding the image and binary closing. This closed image is used to find starch
percentage. Then, starch percentage is calculated by the area and perimeter of cell.
QUALITY REPORT:
The quality report is generated including the type of the sample and date of the report. It
includes a lithe value metrics, Core Type (IC/OC) of the potato tuber, Weight(LIM/S) of the potato
tuber, starch percentage. Also, it includes the type of starch quantity such as high, low or moderate.
The report can be printed or we can generate the QR code to view the report anywhere. This QR
code can be used to identify the potato based on starch percentage. We can usc this QR code in
markets, fields, and stores where potato is being sold. Customers can scan and view the report and
buy the potatoes using this. Also, we can classify the range of potato using this report. We can usc
potatoes with lower starch content for making French fries, and those with high starch content for
mashed potatoes. Also, Potato starch is used in drugs for various purposes.
PROPOSI<:D DDI'M AIKIIIn:CTUtm FOR VIRTUAL STAINING:
The Denoising Difll1sion Probabilistic Models (DDPM) achieve very good performance in
image generation. Compared to GANS, diffusion models are stable and tend to generate large part
of the true data distribution. In DDPMs we use Markov Chain Models which progressively adds
noise to an image. The function q(x, I x,-1), is used to add noise to an image until the image is
essentially pure Gaussian noise at timeT. The cosine scheduler is a learning rate scheduler that
adds noise slower to retain image information for later purpose. But in DDPMs inference speed is
slower than GANs and so it takes longer time to generate images. For that, we can use DDIM
(Denoising Diffusion Implicit Models). DDIMs are a way to speed up the image generation with
a little quality trade-off. To mitigate slower inference speed and to prevent DDPMs from taking
linger time to generate images we. use fewer denoising steps through techniques like I)DIMs
which helps in maintaining the image quality with fewer steps.
CLAIMS
We claim that
Claim I, The system for quality analysis of potato tuber comprising:
-a processing unit:
-to collect cell images of potato tuber either from memory unit or from capturing
unit.
-to process the images and to extract features from them.
- to compare the extracted features with the set uf sorted features is within the
threshold.
-to find if the image is verified or not.
-to recognize the verified image using the deep learning architecture.
-this image is then segmented using segmentation techniques to extract the cell
features from the image.
-to calculate the starch percentage from the segmented image.
-to generate a quality report for this image analysis.
-starch percentage is announceu using voice assistant to make it more enhanced .
. - QR code is generated for the quality report for easily accessible.
-virtual staining is also used here to make this app more cfTective
Claim 2, the system as claimed in Claim I, the extracted features arc related to blurriness,
focus quality, bright spot detection, image texture.
Claim 3, the system as claimed in Claim I, the deep learning architecture is based on MSF[
light-weight architecture along with residual connections.
Claim 4, the system as claimed in Claim I, wherein the processing unit is adapted to
preprocessing techniques to make the background of the image more uniform.
Claim 5, the system as claimed in Claim 2, the verification unit is adapted to usc
Laplacian variance for blurriness detection and also detect bright spots.
Claim 6, the system as claimed in Claim I, the processing unit is adapted to use foldscopc
to capture images where a device is kept on the lens of the foldscope.
Claim 7, the system as claimeu in Claim 6, movable locus ramp is kept to set the focus
and adjustable trace and grip holes arc used to move the foldscopc left/right.