Improving the Representation of CNN Based Features by Autoencoder for a Task of Construction Material Image Classification S. Bunrit, N. Kerdprasop, and K. Kerdprasop Suranaree University of Technology, Thailand Email: sbunrit@sut.ac.th Abstract—Deep learning based model named Convolution Neural Network (CNN) has been extensively employed by diversified applications concerned images or videos data. Because training a specific CNN model for an application task consumes enormous machine resources and need many of the training data, consequently pre-trained models of CNN have been broadly used as the transfer-learning scenario. By the scenario, features had been learned from a pre-trained model by one source task can be proficiently sent further to another specific task in a concept of knowledge transferring. As a result, a task specific can be directly employed such pre-trained features or further train more by setting the pre-trained features as a starting point. Thereby, it takes not much time and can improve the performance from many referenced works. In this work, with a task specific on construction material images classification, we investigate on the transfer learning of GoogleNet and ResNet101 that pre-trained on ImageNet dataset (source task). By applying both of the transfer-learning schemes, they reveal quite satisfied results. The best for GoogleNet, it gets 95.50 percent of the classification accuracy by fine-tuning scheme. Where, for ResNet101, the best is of 95.00 percent by using fixed feature extractor scheme. Nevertheless, after the learning based representation methods are further employed on top of the transferred features, they expose more appeal results. By Autoencoder based representation method reveals the performance can improve more than PCA (Principal Component Analysis) in all cases. Especially, when the fixed feature extractor of ResNet101 is used as the input to Autoencoder, the classified result can be improved up to 97.83%. It can be inferred, just applying Autoencoder on top of the pre-trained transferred features, the performance can be improved by we have no need to fine-tune the complex pre-trained model. Index Terms—Convolution Neural Network (CNN), transfer learning, Autoencoder, construction material, image classification I. INTRODUCTION Since the emerging of deep learning, Convolution Neural Network (CNN) based learning has been extensively employed by diversified applications. Especially, for the tasks concerned images or videos data. Due to the constructing and learning of a specific CNN Manuscript received April 17, 2020; revised September 23, 2020. model for an application task consumes enormous machine resources and need many of the training data, consequently pre-trained models of CNN have been published and appreciation by many application domains. The features had been learned from a pre-trained model by one source task can be proficiently sent further to another specific task in a concept of transfer learning. By transfer learning, a task specific can be directly employed such pre-trained transfer features or further train more by setting the pre-trained transfer features as a starting point. Thereby, it takes not much time and can improve the performance from many referenced works. Transfer learning of CNN model can be applied by two schemes, which are fixed feature extractor and fine-tuning. Fixed feature extractor directly transfers pre-trained features to a task specific by just project (activate) the task specified data to such features. Another one popular scheme is fine-tuning. It means the pre-trained transfer features from a source task are fine-tuned to a task specific by training more with a task specific dataset. The result features after retrain are then utilize. Naturally, fixed featured extractor can be process faster than fine-tuning, especially when the pre-trained model is very deep. The deeper of the model, the longer of the fine-tune process. In addition, fine-tuning process need to set many of hyper-parameters. Searching for such suitable and optimal hyper-parameters also take much of time and complex. In this research, aim at looking for the best performance in construction material images classification task, the novel suitable approaches are then explored. Previous works concerned construction material image classifications were studied based on hand-designed features, of which the outstanding algorithms in image analysis were applied to extract the features and then some classifiers were selected to classify such features. Therefore, the classification accuracy depends on manual selection of the feature-extracted algorithm. In our study, the state of the art approach based on the transfer learning of CNN pre-trained models/architectures is investigated. A set of construction material images is act as a task specific dataset in the transfer-learning scenario. The two selected architectures are GoogleNet [1] and ResNet101 [2] pre-trained on a source task of ImageNet dataset. These two architectures are differences both in deep and in detailed layers. Journal of Advances in Information Technology Vol. 11, No. 4, November 2020 © 2020 J. Adv. Inf. Technol. 192 doi: 10.12720/jait.11.4.192-199