Image sentiment classification

2017 · Machine Learning

Github Repo

Problem description:

Use a CNN model to classify 7 types of sentiment, which are anger, fear, disgust, surprise, happiness, sadness, and neutral

CNN model

• Conv2D: This layer creates a convolution kernel that is convolved with the layer input to produce a tensor of outputs
• Activation('relu'): Applies the rectified linear unit activation function $max(x, 0)$
• ZeroPadding2D: Zero-padding layer for 2D input
• MaxPooling2D: Max pooling operation for 2D spatial data

Training and Testing Accuracy in a CNN model

Compare the accuracy between a DNN and a CNN model

• A CNN model has better performance in feature detection of images than a DNN model.

Confusion matrix between each class

• “Sad” is easily misunderstood as “neutral”
• “Digest” is easily misunderstood as “angry”
• “Fear” is easily misunderstood as “sad”

Result

• Achieved 161/366 (Top $44\%$) rank in the Kaggle competition