Image Aesthetic Assessment using Deep Learning based on the photography rules such as Rule of Thirds, Depth of Field, and Color Contrast. This is the second and final part of my final year BE project. The first part is contained here: https://github.com/ananyapal/Image-Aesthetics-Handcrafted
Although our dataset is not publicly available, the code for the training and testing are available and any image dataset can be used for training.
- Coordinated a team of 4 and collected a dataset of 6000+ images satisfying 3 high-level photography rules as mentioned above.
- Preprocessed images into grayscale, resized and scaled into 128*128 images, and passed through 5 convolutional layers with Max Pooling and 2 fully connected layers (ReLU) and one output layer (Sigmoid).
- The Convolutional Neural Network (CNN) finally classified images as aesthetically pleasing or non-pleasing.
- Achieved an accuracy of 68% through the Deep Learning model on the assembled dataset compared to traditional and Machine Learning (SVM) methods which only scored 40%.
- For more details, take a look at our work that was published in the IJRASET journal, also named as 'author(1).pdf'.