An Approach to Enhance Image Recognition Using Improved Triplet-Loss and Reinforcement Learning and Active Learning

Image recognition is increasingly more and more important in modern life. In fact, it is used not only in the medical field, but also in autonomous systems such as robotics and self-driving vehicles. Therefore, the development of a high-accuracy recognition method is becoming increasingly necessary. Today, one of those methods is to represent embedding vectors more clearly in the space, followed by an anchor vector, such as Triplet-loss. These embedding vectors are then fed into many deep learning models for the recognition phase. These vectors help those deep learning models recognise them with higher accuracy. However, embedding vector representation methods like Triplet-loss still have limitations where their representation vectors are not sufficiently clearly represented in a high-dimensional space. Therefore, in this paper, we propose a novel framework that integrates a new method for embedding representation based on Triplet-loss, namely Triplet-LINEmbedding, to improve the representative embedding vectors; from there, deep learning models can leverage their recognition accuracy. Furthermore, we also combine reinforcement learning and active learning into the framework to enhance the ability of deep learning models to recognise images. The result of our experiments is evaluated using 5-fold cross-validation on two different datasets, which shows that our proposed method improves recognition performance on these datasets.