Advancing Deep Learning Methods for Long-tailed Image Recognition

One of the most important factors for the success of any deep learning model is the distribution of data. The effectiveness of deep learning models is significantly affected by the distribution of data within the training data set. Deep learning algorithms have already achieved outstanding performance in terms of accuracy and efficiency on the balanced distribution of data; however, if the data distribution in the dataset is imbalanced, i.e., long-tailed, it results in a biased classifier causing the degraded performance of the deep learning model. Long-tail distribution-based image recognition has attracted much attention from researchers in the field of computer vision since the last decade. Different methods have been proposed to address the long-tail distribution-based image recognition problem in the existing literature, however, there is still room of improvement. So, this research is aimed to address the problem of long-tail distribution-based image classification for deep learning models. In particular, this research unfolds its investigation into how to mitigate the long-tailed data distribution challenges against the biased classification from the following three aspects.Firstly, this thesis argues that the diversity of image features, inter-class distance and intra-class distance play a vital role in image recognition and enables the classifier to classify the images correctly. Accurate classification requires well-separated classes (large inter-class distance) and compact sample distributions within each class (small intra-class distance). In long-tailed settings, tail classes contain few and less diverse samples, which leads to larger intra-class distance and consequently reduces their discriminability. So, this thesis has proposed and implemented a novel method of enriching the feature space of tail-class samples, i.e., incorporating the variance of angles from the head-classes to the tail-classes with supervised contrastive learning. Performing experiments on different long-tailed benchmark datasets shows that this method has also outperformed many of the existing baseline methods.Secondly, this thesis introduces a novel approach that augments the traditional feature scaling mechanism by incorporating adaptive margin constraints into the scaling factor itself, thereby enabling dynamic expansion of the feature space throughout training. In conventional long-tailed recognition, fixed or uniformly applied scaling often fails to sufficiently separate minority class features from those of the majority, especially in high-dimensional embedding spaces where class boundaries tend to collapse under imbalance. This thesis addresses this by embedding class-dependent adaptive margins within the scaling formulation, allowing the model to not only compact intra-class features but also actively push inter-class boundaries outward in a controlled manner. This adaptive margin is updated jointly with the network parameters in an end-to-end optimization process. Empirical evaluations demonstrate that this adaptive margin integration consistently boosts Top-1 accuracy, F1-score, and AUC across multiple long-tailed benchmarks, confirming its effectiveness to existing classification strategies.Finally, this thesis addresses the joint presence of long-tailed distributions and noisy labels. The proposed framework, i.e., Hybrid Distribution-aware Graph and Prototypical Contrastive Learning (HDG–PCL) departs from purely batch-wise remedies by maintaining a global memory bank that stores high-confidence samples identified via the consensus of a conventional classifier and a balanced classifier throughout training. These global prototype and local graph (hybrid) objectives are optimised alongside supervised and mixup-based semi-supervised losses, culminating in balanced, noise-tolerant representations. Extensive experiments on long-tailed datasets demonstrate that HDG–PCL sets a new state of the art under simultaneous presence of imbalance and noise factors.In summary, this thesis investigates, designs, and implements advanced methodologies to mitigate the adverse effects of long-tailed data distributions on deep learning performance. Through the introduction of three novel contributions, i.e., Supervised Angular Contrastive Learning with Balancing Classifier (SACL-BC), Optimized Adaptive Feature Compression (OAFC) to enhance feature space geometry, and Hybrid Distribution-aware Graph and Prototypical Contrastive Learning (HDG–PCL) that synergistically combines global and local representations, this work advances the state of the art in achieving higher image classification accuracy under severe class imbalance conditions. Despite the demonstrated effectiveness of the three proposed contributions, each method carries certain limitations. First, the feature-space enrichment strategy (SACL-BC) requires the use of a manually selected threshold to separate head and tail classes for different imbalance ratios, which may limit its adaptability across datasets with varying distributions. Second, the adaptive-margin scaling approach (OAFC) introduces additional hyperparameters and relies on stable margin updates during training; hence, its performance may be sensitive to margin initialization and optimization settings, potentially requiring careful tuning for different architectures or datasets. Third, although the HDG–PCL framework effectively handles the joint presence of imbalance and noisy labels, maintaining a global memory bank increases computational and memory overhead, and the method’s reliance on consensus-based pseudo-labeling may still be affected when both classifiers exhibit high uncertainty in extremely noisy scenarios.