The findings demonstrate that introducing prior structure, through strategic sparse supervision, pre-trained multimodal knowledge, or transferable representations, consistently compensates for limited labeled data across diverse vision tasks.
Abstract
Deep learning has enabled the wide-spread deployment of computer vision beyond laboratory settings. Consequently, computer vision has become the main bridge for intelligent systems observing the real, physical world. Yet, a fundamental barrier persists: the substantial volume of precisely labeled training data required for effective supervised learning. Especially in industrial settings, where public datasets rarely match domain-specific requirements, the combined effort of data capture, cleansing, and annotation can account for up to 80% of project costs, constraining deployment agility and accessibility. This data hunger manifests distinctly across tasks. For example, semantic segmentation can demand pixel-perfect masks, while anomaly detection often requires extensive normal examples and, in some cases, defective examples, which are rare by nature. Some failure cases might even be entirely unknown while the inspection pipeline is being built. This dissertation develops and evaluates methodologies that reduce dependency on extensively labeled datasets. The central research objective is to demonstrate that strategic approaches, leveraging sparse supervision, pre-trained multimodal knowledge, and transferable embeddings, can achieve competitive performance while substantially reducing annotation requirements, thereby making deep learning more practical for resource-constrained deployment.
Three empirical investigations address complementary aspects of data efficiency. First, a “Lazy Labels” technique enables U-Net architectures to learn semantic segmentation from sparse point or scribble annotations by zero-masking unlabeled pixels during loss computation, requiring modifications only to label tensor handling. Second, Multimodal Large Language Models (MLLMs), including GPT-4V and Qwen-VL-Chat, are evaluated for visual anomaly detection through a training-free Visual Question Answering (VQA) framework, using natural language prompts with zero to a few reference images. Third, a comparative study contrasts pre-trained foundation model embeddings (MegaDescriptor) against task-specific training for visual re-identification, examining both frozen and fine-tuned transfer learning under standard and one-shot conditions. The sparse segmentation experiments demonstrate that models trained with point annotations covering ≈1% of pixels achieve performance within 1− −3 F1-score percentage points of fully supervised baselines, while producing spatially coherent dense predictions. For anomaly detection, GPT-4V achieves F1 = 0.92 in a one-shot configuration on MVTec AD without task-specific training, approaching specialized detector performance. In re-identification, fine-tuned pre-trained embeddings reach 91.6% top-1 accuracy under full supervision and notably outperform from-scratch training in data-limited one-shot scenarios, validating their superior data efficiency.
The findings demonstrate that introducing prior structure, through strategic sparse supervision, pre-trained multimodal knowledge, or transferable representations, consistently compensates for limited labeled data across diverse vision tasks. While each approach operates within specific constraints, including sparse label distribution, sensitivity to foundational capabilities, and domain similarity, they collectively demonstrate viable pathways for reducing the data bottleneck. The work contributes observations and validated techniques for practitioners: zero-masking for direct sparse supervision, prompt-based MLLM deployment for rapid baselines, and transfer learning strategies optimized for data scarcity. These methodologies advance the practical deployment of computer vision in contexts where extensive annotation remains prohibitive.
Tremendous numbers of images and videos are generated throughout the life cycle of construction projects, offering rich visual data that, if analysed effectively, can support automation in progress monitoring, safety management, quality control, and productivity tracking. Within this context, image classification plays a foundational role: it is the supervised task of mapping an image to one of K semantic categories by learning from labelled examples, requiring the extraction of discriminative visual structure and the production of calibrated class predictions. Despite major advances from handcrafted feature pipelines to Convolutional Neural Networks (CNNs), fair comparison across methods remains challenging due to inconsistent preprocessing, dataset splits, training budgets, and reporting standards. This paper introduces a unified and fully reproducible benchmarking framework that standardises preprocessing, enforces fixed train–test splits, aligns training configurations, and evaluates models using a consistent multi-metric suite. Under identical experimental conditions, we benchmark four paradigm families: classical machine learning (ML) models, handcrafted feature-based methods (Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), Bag-of-Visual-Words (BoVW) [Formula: see text] Support Vector Machine (SVM)), and CNNs (LeNet-5, ResNet-18), across three datasets (MNIST, Fashion-MNIST, CIFAR-10). We formalise image classification mathematically as a supervised learning problem and assess performance using accuracy, macro-averaged [Formula: see text]1-score, and macro-averaged ROC–AUC, complemented by learning-curve and robustness analyses. Results show that ResNet-18 achieves the strongest performance on all datasets (98.95%, 90.27%, 72.90%), SVM with a Radial Basis Function (RBF) kernel remains competitive on grayscale benchmarks, and hand-crafted pipelines degrade sharply on CIFAR-10. ROC–AUC analysis further highlights calibration–accuracy discrepancies, such as Naïve Bayes achieving high AUC but modest accuracy on MNIST. These findings emphasise the importance of task-aware model selection and demonstrate that compute-aligned, multi-metric benchmarking is essential for fair and reproducible cross-paradigm evaluation. All CNN results are reported under a 50-epoch training budget, which represents a lower bound on converged performance; future extensions will incorporate domain-specific datasets and pre-trained architectures to broaden practical applicability.
Yasir Hamid, N. Ababneh, A. Abasi et al.· Journal of Information &...· 0 citations
Experimental findings show that pretrained models outperform those trained from scratch in terms of accuracy, convergence speed, and robustness, and a unified framework is proposed to integrate both processes in a deep learning pipeline.
Kwame Nkosi· International Journal of App...· 0 citations
The rapid growth of artificial intelligence and deep learning has revolutionized diverse domains, including computer vision where transfer learning through pre-trained models has become a fundamental technique for enhancing performance of models and reducing computational costs. Although numerous pre-trained deep learning models have been widely adopted, their effectiveness varies significantly across application domains and datasets. The lack of comprehensive comparative studies makes it challenging for researchers to identify the most suitable models for achieving optimal performance in specific tasks. This study systematically compares seven pre-trained feature extractors across three architectural families, convolutional neural networks (CNNs), Vision Transformers (ViTs), and self-supervised models to provide practical guidance on model selection for downstream deep learning tasks. These models were evaluated on five benchmark datasets. Features were extracted with frozen weights and evaluated using linear probing, k-nearest neighbor retrieval, and 5-shot classification. ConvNeXt-B achieved the highest mean linear probe accuracy (90.40%), while DINOv2-S produced the best feature geometry for retrieval tasks (87.64%). CLIP-ViT-B/32 demonstrated the strongest cross-domain transfer, leading on texture recognition and satellite imagery few-shot classification. Older CNN architectures lagged significantly behind modern models by approximately 7%. Overall, this study simultaneously evaluates multiple feature extractors across multiple visual domains and evaluation protocols.
Rafeek Sibrikhan, M. Mufassirin· Sri Lankan Journal of Techno...· 0 citations
Although SDG methods improve performance under highly domain distinguishable stylized shifts, they exhibit limited robustness to background, correlation and corruption shifts on larger datasets, and increased shape bias does not consistently yield enhanced OOD performance, thereby underscoring the need for further research into developing more resilient and generalizable models.
K. Imbulgoda, Ruwan Tennakoon, W. Chuah et al.· International Journal of Com...· 0 citations
Robotic perception pipelines increasingly rely on large vision backbones deployed on SWaP-constrained edge platforms, making post-training quantization (PTQ) attractive for real-time inference. However, while PTQ often preserves clean in-distribution accuracy, we show that it can substantially degrade reliability under deployment-relevant distribution shifts (e.g., sensor noise, severe weather, and novel operating environments), creating a Quantization-Induced Robustness Gap. Across foundational vision benchmarks (ImageNet-C and PACS), 4-bit PTQ models exhibit pronounced robustness degradation despite negligible ID accuracy loss. To address this, we propose Recti-Q, a lightweight feature-space rectification framework that freezes the quantized backbone and trains a small classifier-head LoRA adapter using only source data. Recti-Q is architecture-agnostic across CNNs and Transformers, supports efficient teacher-free training, and recovers a significant portion of the lost robustness, in some cases matching or exceeding FP32 performance. At less than 1% parameter overhead (as small as 6 KB), Recti-Q preserves over 99% of PTQ memory savings, adds negligible compute, and enables low-bandwidth Over-The-Air (OTA) resilience patching for deployed robotic fleets operating in unpredictable physical environments.
Despite the growing need for object-detection systems, companies, governments, industries, and organisations face significant challenges in implementing them to address real-world problems. Object detection supports applications ranging from quality control and inventory management to robotics, surveillance, autonomous systems, and defect detection. Advances in deep learning have enabled notable progress through architectures such as convolutional neural networks (CNNs). This study examined the performance of two deep-learning architectures for detecting household items under controlled experimental conditions. A dataset of 2,100 images, comprising three balanced classes—bottles, boxes, and cups - was annotated and used in the experiment. The results showed that the fine-tuned VGG16 model consistently outperformed the plain CNN across the principal performance measures. Intersection over Union (IoU) increased from 57.11% for the plain CNN to 98.00% for the fine-tuned model, while loss decreased from 6.40% to 4.21%. The fine-tuned model also produced a smoother and more stable learning curve, indicating more consistent detection performance. However, this improvement required a longer training time: approximately 14,400 seconds for the fine-tuned model compared with 12,240 seconds for the plain CNN. The comparison was conducted using the same dataset, training duration, and evaluation framework for both architectures. These findings indicate a trade-off between computational efficiency and predictive performance under the experimental conditions used.
Anas Tukur Balarabe, Najib Hassan Adamu, Mahmood Umar et al.· Asian Journal of Research in...· 0 citations