image classification with a simple nn

notes link - https://miro.com/app/board/uXjVIPolTio=/?share_link_id=225966248894 colab - https://colab.research.google.com/drive/15D91NChzSrEM2kTr9E8GI6l1FGvu6men?usp=sharing

architecture of our nn

• here we build a shallow nn with no activation fn to see how well it can perform in image classification

as a general concept, the model should atleast perform better than 20% accuracy as a random guesser would still have 20% acc because there are 5 classes

• classes - daisies, dandelions, roses, sunflowers, tulips
• each pic is jpeg and each category has 1000 images
• image processing : load → jpeg to rgb → scale 0-1 → resize 224 x 224 x 3
• batch dim img → 16 x 224 x 224 x 3, label → 16 (labels are turned into numbers for each class 0-4)
• architecture - flattening → dense FC → softmax

results of the experiment

how to improve?

• use deep nn with more hidden layers
• use non-linear activation
• use convolution operation

Links:

202606051832