TOCNet: A target occlusion contrast network for license plate detection in waterways

Abstract

With the continuous development of waterway trade, intelligent waterway supervision is gradually becoming an important way to maintain waterway safety. In this field, ship license plate (SLP) detection is importance for intelligent waterway supervision. However, in practical surveillance scenarios, detecting ship license plates (SLPs) is challenging because of small size, and text areas in complex backgrounds can also be mistakenly identified as SLPs. To solve these problems, we propose a Target Occlusion Contrast Network (TOCNet) for SLP detection. Firstly, to mitigate the effects of complex backgrounds, a novel Target Occlusion Contrast Learning (TOCL) approach is proposed, which is used to widen the decision boundary of the model and reduce false alarms. Secondly, to solve the problem that small targets are difficult to be detected, we design a Similarity Fusion Module (SFM), which effectively improves the detection performance of the model for small targets by similarity weight. Finally, to further improve the localization accuracy, we propose the task-oriented decoupled head (TDH), which facilitates the localization accuracy through selective fusion. Experimental results on both the SLPD dataset and the UFPR-ALPR dataset show that our proposed modules effectively improve the detection performance of the model, and the proposed TOCNet model exhibits a better detection performance than the other existing state-of-the-art (SOTA) models. Our code and dataset are publicly available at https://github.com/ssyuyi/tocnet.