sindhri / object_classification

using MLP and CNN for object classification

Object Recognition

Source: DPhi challenger 31 https://dphi.tech/practice/challenge/31

Objective: classify images into 10 classes of objects:

• Ship
• Aeroplane
• Automobile
• Bird
• Cat
• Deer
• Dog
• Frog
• Horse
• Truck

Training data is the cifar10 library: size 50,000. Image size 32x32.

Simple EDA

Data preprocess

• normalization

build models

3.1 simple MLP

Did not yield satisfying results

3.2 deeper MLP

Did not yield satisfying results

3.3, 3.4 deeper MLP with more parameters

Did not yield satisfying results

3.5, simple CNN

test accuracy is 69%, consistent with validation accuracy

3.6, larger CNN

after 50 epochs (18 hours), the validation accuracy is still only 78%, test accuracy was the same