Weakly Supervised Logo Detection Approach

The previous chapters discuss Deep Learning-based feature extractor frameworks and challenges related to detection tasks, with a specific focus on logo recognition. Most existing logo detection methods often rely on precise object-level bounding box (position bounding box) annotations, that poses substantial challenges in practical settings due to the labor-intensive nature of object-level annotations. To address this issue, this chapter presents a novel weakly supervised logo detection algorithm that enables effective logo recognition without necessitating detailed bounding box annotations. We begin by exploring approach to weakly supervised logo recognition that utilizes only image-level annotations. We explore the integration of attention mechanisms and a feature extraction network using dilated convolutions, aimed at compensating for the lack of precise object localization typically provided by bounding box annotations. In a weakly supervised training scheme, we lack guidance on locating object positions as bounding box annotations are not available during training. The primary goal is to boost performance by adeptly utilizing image-level labeled data. To enhance logo image classification and localization, the chapter introduces the application of attention-based mechanisms with Convolutional Neural Networks (CNNs). These mechanisms are designed to emphasize critical attributes and information, assigning greater significance to the spatial and semantic aspects of objects. This scheme relies on image-level annotations and demonstrates substantial scalability and adaptability for widespread real-world deployment.