M Sc

The research work entitled “Statistical assessment of banana ripening using
smartphone-based images” was carried out at College of Agriculture, Vellayani during
the period 2019 to 2021. The objectives were the development of suitable model to
establish the relationship between Total Soluble Solids (TSS) and L*(lightness),
a*(green-red ratios), b*(blue-yellow ratios) values and for prediction of TSS values
using L*, a*, b* values. Development of a protocol for accurate data collection to assess
TSS content in Banana using smart-phone-based images. Good quality Nendran variety
with only minor shape and peel colour flaws were obtained from a nearest field
randomly chosen for the study. Each time 3 hands at the ripening stage 1 (green) with
10 fingers by hand were collected. The fruits were stored in a normal day/ night cycle.
Bananas were taken randomly from each hand and their color changes and development
of brown spots were measured daily during 10-12 days. Banana samples were placed
on the table covered with a non-reflecting white paper as a background of the image.
For white light illumination, two of 36 W fluorescent lamps were fixed at ceiling above
the experiment setup. Three smartphones were used for image acquisition. Smart
phones were placed at a distance of 20 cm above the banana. Samples of banana were
blended using a fruit juicer. The TSS were determined using a digital refractometer. For
the images obtained, RGB and L*a*b* were extracted using ImageJ software. The
observations on TSS, R, G, B, L*, a*, b* were used for fitting regression models after
splitting the data into train (80%) and test (20%) sets.
When linear model was fitted between TSS and R, G, B values for all the three
devices, each of the independent variables were found to be significant. Adjusted Rsquared values obtained were 0.80, 0.80, and 0.84 for the three devices. It means about
80% of the variation in the TSS was explained by R, G, B values. For the predicted
values of TSS R-squared values were 0.84, 0.90, and 0.95. Hence linear model was
found to be better fit for predicting TSS. Since RGB color model is device dependent
model, it may not always represent the same colour on different devices. But in case of
CIE L*a*b*, it is device independent and shadows and areas of glossiness on the object
surface had less impact. Therefore, linear model was fitted between TSS and L*, a*, b*
values. Adjusted R-squared values obtained were 0.78, 0.81, and 0.85 for the three
126
devices. For the predicted TSS values R-squared values were 0.84, 0.76, and 0.95.
Therefore, linear model between TSS and RGB model found to predict TSS much
accurately than L*a*b* color space when prediction accuracy was compared.
On visualization of data, TSS and L*a*b* found to have non-linear relationship
for all the devices. When spline regression was fitted between TSS and L*, a*, b* values
R-Squared obtained were 0.91, 0.90, and 0.89, which was higher compared to Rsquared values for linear model. Also, deviance explained by the models were 92%,
92.3%, and 90.7% for corresponding device 1,2 and 3. Therefore, spline regression
found to be better model for TSS and L*, a*, b* data and for prediction of TSS values.
Protocol for accurate data collection was developed with modification in the
procedure performed. Possibility of Deep learning was explored in the study using
CNN. Convolutional neural network (CNN) was developed using 3 categories Raw
(TSS 4-10), Medium (TSS 11-17) and Ripe (TSS 18-32) with 30 samples each. 25
images from each category were taken as training set and 5 were taken as test set. 100
epochs were performed to mitigate overfitting and to increase the generalization
capacity of the neural network. Model evaluation of training set gave an accuracy of
84% with loss value 0.45. For the training set, all 25 from ripe category were able to
identify into that particular category. In case of raw 24 were identified as raw with 1
identified as medium. For medium 14 were identified as medium,3 identified as ripe
and 8 identified as raw. Model evaluation of test set provided 73% accuracy with 0.81
loss. The model successfully classified 5 ripe bananas, 4 raw bananas (1 classified as
medium) and 2 medium bananas (3 classified as raw).
The results of the research work to identify the best fitting model concluded that
RGB model found to predict TSS much accurately than L*a*b* color space when linear
regression model was fitted and spline regression model was found to be the best fit for
L*, a*, b* and TSS values, R-squared values were much higher with a good percentage
of variation explained. The CNN developed classified images into raw, medium, and
ripe with approximate accuracy of 74%. Therefore, CNN can be used to predict range
of TSS in no time, if a large number of images are uploaded into this model. The CNN
can be optimized further with higher number (atleast 10,000 samples) of samples to
improve the efficiency of classification.

English