Leveraging local diversity in weakly supervised fine-grained image classification

Аннотация

Fine-grained image classification (FGIC) presents a challenge due to the high similarity between visually similar categories, requiring precise feature extraction. Traditional supervised methods demand large amounts of labeled data, which is often unavailable. This paper introduces a local diversity-guided weakly supervised method to address these challenges by leveraging weakly annotated data, thus alleviating the dependence on fully labeled datasets. The approach focuses on fine-grained object recognition by combining local diversity features with weak labels, which significantly improves classification performance. Experimental results on several benchmark datasets demonstrate the method's effectiveness in enhancing the precision of fine-grained classification tasks. Keywords: Fine-grained image classification, weakly supervised learning, local diversity guidance, deep learning, multi-instance learning, image-level labels, discriminative features, attention mechanisms, benchmark datasets, classification accuracy. labels. Other techniques, such as part-based models Introduction: Fine-grained image classification (FGIC) and attention mechanisms, have also been widely aims to differentiate between visually similar sub- studied, focusing on locating and identifying categories of objects, such as distinguishing between discriminative regions in images. species of birds, types of flowers, or model variants of cars. These tasks require not only high accuracy but Recent advancements in local feature extraction and also the ability to capture subtle differences among self-supervised learning have also shown promising similar-looking classes. Traditional methods for FGIC results. However, challenges remain, particularly in often rely on fully labeled datasets, which are fine-grained classification, where subtle visual expensive and time-consuming to create. differences are crucial. Furthermore, most existing methods still rely heavily on large labeled datasets or Weakly supervised learning (WSL) methods, in are constrained by the availability of high-quality contrast, rely on limited or weak annotations—such as annotations, which limits their applicability in real- image-level labels—instead of pixel-wise or object- world scenarios. This study addresses these gaps by level annotations. This paper presents a novel local proposing a new method that efficiently uses weak diversity-guided weakly supervised fine-grained image annotations and improves classification accuracy classification method, designed to enhance FGIC by through local diversity guidance. focusing on local regions of images. The method aims to leverage the power of weak annotations while METHODOLOGY minimizing the reliance on large, fully annotated 1. Weakly Supervised Learning Framework datasets. We begin by utilizing a weakly supervised learning Related Work framework, where only image-level labels are provided Over the years, several approaches have been during training. This alleviates the need for detailed proposed for fine-grained image classification, annotations such as bounding boxes or part-level especially using weakly supervised learning techniques. labeling, which are often costly to obtain. The goal is to Methods such as multi-instance learning (MIL) have leverage the image-level label to guide the learning been explored to train classifiers based on image-level process for fine-grained classification tasks. American Journal of Applied Science and Technology 1 https://theusajournals.com/index.php/ajast American Journal of Applied Science and Technology (ISSN: 2771-2745) 2. Local Diversity-Guided Feature Extraction (mAP), to assess the classification performance. Our method consistently outperforms existing weakly Our method introduces a local diversity module, which supervised approaches, particularly on datasets with focuses on identifying and utilizing diverse regions highly similar categories. within the image to capture discriminative features. The idea is to identify multiple regions that contribute 3. Comparison with State-of-the-Art Methods uniquely to the classification task. By learning from Our approach shows a significant improvement in these diverse local regions, the model can better classification accuracy compared to existing methods. distinguish between visually similar categories, even In particular, the integration of the local diversity when only weak annotations are available. guidance leads to better generalization on unseen The local diversity is achieved by considering features examples, demonstrating the effectiveness of our extracted from various regions of the image. We use a method in fine-grained image classification tasks. convolutional neural network (CNN) to extract high- RESULTS AND DISCUSSION level features, and then employ a clustering technique The experimental results demonstrate that the local to group regions that exhibit diverse characteristics. diversity-guided weakly supervised method These diver