Quad-Channel Contrastive Prototype Networks for Open-Set Recognition in Domain-Specific Tasks

Abstract

A traditional deep neural network-based classifier assumes that only training classes appear during testing in closed-world settings. In most real-world applications, an open-set environment is more realistic than a conventional approach where unseen classes are potentially present during the model's lifetime. Open-set recognition (OSR) provides the model with the capability to address this issue by reducing open-set risk, in which unknown classes could be recognized as known classes. Unfortunately, many proposed open-set techniques evaluate performance using ''toy'' datasets and do not consider transfer learning, which has become common practice in deriving a strong performance from deep learning models. We propose a quad-channel contrastive prototype network (QC-CPN) using quad-channel views of the input with contrastive prototype loss for real-world applications. These open-set techniques also require the tuning of new hyperparameters to justify their performance, so we first employ evolutionary simulated annealing (EvoSA) to find good hyperparameters and evaluate their performance with our proposed approach. The comparison results show that QC-CPN effectively outperforms other state-of-the-art techniques in rejecting unseen classes in a domain-specific dataset using the same backbone (MNetV3-Large) and could become a strong baseline for future study.

Vehicle Make and Model Recognition as an Open-Set Recognition Problem and New Class Discovery

“…OSR led to extensive research that mostly focused on more effectively limiting the open-space risk [15][16][17][18], and little research was developed around efficiently performing open-set recognition and simultaneously discovering new classes hidden in the rejected instances. Some of the proposed solutions employed incremental learning [19], transfer learning [20,21], or clustering [22,23].…”

“…Numerous studies have been conducted to minimize the risk of open sets and more effectively reject objects of unknown classes [15][16][17][18], which is the main goal of OSR. However, in a more desirable context, an OSR should go further and discover the unknown classes hidden inside the rejected objects.…”

“…Another proposal that may be related to our work was presented in [18], where the OSR problem was addressed within a transfer learning approach using contrastive learning to model the data. They also highlighted the importance of developing and testing OSR solutions with domain-specific databases to test their efficiency in dealing with real-world applications.…”

Towards open domain-specific recognition using Quad-Channel Self-Attention Reciprocal Point Learning and Autoencoder

“…To address these challenges, we propose the utilization of Reciprocal Point Learning (RPL) and the autoencoder. We extend the RPL by diversifying the input of the deep neural networks through geometric rotation inspired by our previous work [1] and process these input representations using a self-attention mechanism. Autoencoder was utilized to distinguish between unknown ID and OOD as it contains the information of the reconstructed image, including the background that is useful to determine their unknown distribution, e.g., food-related images often feature hands or plates, while the vehicle commonly found in the garages or on the roads.…”