Combination of Feature Engineering and Feature Learning Approaches for Classification on Visual Complexity Images

Indah Agustien Siradjuddin, Afni Sakinah, Mochammad Kautsar Sophan
2021 International Journal of Innovative Computing, Information and Control  
The feature engineering approach determines features using expert knowledge about the data input and feature extraction method. Therefore, this approach obtains features that are representing the input for the classification problem. The feature learning approach relies on the learning process that maps the raw data input and its target. The obtained features are the representation of data input based on its target. This paper presents feature engineering and feature learning for image
more » ... ation based on its image complexity. Structure, noise, diversity, and number of grouped regions result from feature engineering of an image. The convolution process between input images and the Laplacian filter obtained structure and noise features. Diversity is the number of detected corner pixels, and the connected component labeling algorithm extracts the number of grouped regions. Meanwhile, the feature learning results from the convolutional layers of the convolutional neural networks architecture. The classification layers of the proposed convolutional neural networks architecture used the combination of these features. For the experiments, we built and trained three proposed models based on the input features. First, we used the feature engineering approach. Second, the feature learning approach and the third model used the combination of feature engineering and feature learning. The experiments were conducted on thousands of collected images. The results show that the combination features achieved the highest average accuracy, up to 75.15%, since feature engineering emphasizes feature learning for the classification.
doi:10.24507/ijicic.17.03.991 fatcat:ik6fk4dwlncqnpkcrznctgeuxe