/g/ - Technology

I trained a cnn on this dataset

https://data.mendeley.com/datasets/b6fftwbr2v/1

[Blocked URL: https://www.sciencedirect.com/science/article/pii/S2352340921009616]

then i did feature extraction and did k means clustering and there were 4 clusters i made, now thing is, in this sample images in picrel it shows fruits in cluster 1 and 2 which is mid ripe and ripe, 0 and 3 are unripe and overripe, now that i tried this model on different images from internet, it only predicts unripe or overripe although it predicts overripe and unripe quite perfectly but never does anything for mid ripe and unripe.

>>824(OP)

Did you use pca or fca?

>>827

since the database has info about class ( unripe ovveripe etc) you should use fca it will seperate classes more which should help distinguish more.

>>828

also try xgboost on fca data , xg boost is really good . :)

thats one method test it .

another might be doing regression. on fca dataset.

>>827

another way except all this is to redefine model ,

instead of model predicting ripe unripe directly , you can divide this into two part , one model predicts what kind of fruit that is ( which has high accuracy , easy to make ) and with the output of that you decide whether that fruit is ripe unripe etc now for this part you can make model for each class using loops or whatever ( use dataset of that class only). this will be better than all other as this will work as a pigeonhole , we will have selective data to work with which can help us distinguish more.

>>832

>>830

>>833

model doesnt predict cluster directly, I have first taken labeled data and trained a cnn on it to predict quality of fruit( good, bad and mixed) then type of fruit, so like i had 3 classes bad good and mixed and each had 6 types of fruits. now i trained cnn on it then did feature extraction from dense layer and after that i have taken it as a complete raw dataset and applied k means on those extracted features. wait, i will post pca for these clusters. there seems to something wrong with mid ripe and ripe cluster, also i am just gonna change clusters to 3 instead of 4 and mid-ripe and ripe will keep overlapping.

>>834

also this is how my dataset looks, there seems to be imbalance too in my dataset, now i have noticed, like pomegranate of good quality are 6000 alone and rest 1k. when i showed this to my prof, she said, what was the point of applying clustering in the end after feature extraction when dataset is labeled, idk if she misunderstood or i am doing something wrong.