Robust and Resource-Efficient Computer Vision Pipelines for Open-World Industrial Applications

Deploying deep learning frameworks for object instance segmentation in unstructured industrial environments is constrained by critical limitations across data curation, model optimization, and edge hardware verification. While visual computing models achieve high accuracy in controlled laboratory settings, operational environments introduce severe noise distributions including clutter, dynamic object scales, varying lighting conditions, and unforeseen physical anomalies. This dissertation presents a unified perception stack designed to mitigate these core constraints, validated on complex industrial manufacturing and agricultural sorting tasks. To automate data engineering lifecycles, we first introduce the Fast Object Segmentation and Tracking Tool (FOST). By coupling interactive zero-shot click-segmentation with temporal Lucas-Kanade optical flow, FOST reduces tracking annotation overhead to fractions of a second per frame. This pipeline enabled the curation of SteelDS, an open-source, high-resolution video benchmark dataset containing 24,297 labeled frames that map multi-scale material fractions under diverse clutter densities. To satisfy live edge-processing latency requirements, we propose Instance-Based Importance Scores (IBIS), a training-aware gradient sparsification algorithm. By conditioning weight updates on target object instances, IBIS achieves up to a 95% parameter compression rate in convolutional and transformer backbones while maintaining comparable object segmentation and classification performances. Furthermore, a systematic evaluation of a number of architecture variations of the sparsified models under seven distinct environmental corruptions (including motion blur, atmospheric dust, and spatter) reveals that unstructured sparse networks function as robust spatial filters, outperforming dense baselines under domain shifts. Finally, to reduce the substantial manual overhead of generating pixel-precise labeled training targets, which remains a core challenge in supervised learning methods, we explore a novel approach to unsupervised discovery. We present PASTA (Vision Transformer Patch Aggregation for Weakly Supervised Target and Anomaly Segmentation) for zero-shot target and anomaly segmentation. By evaluating token distribution contrasts against nominal references, this framework demonstrates domain-agnostic generalization across metal recycling tracks and autonomous agricultural crop tracking. The core architectural and methodological contributions presented in this dissertation are substantiated by two published peer-reviewed conference papers, two journal manuscripts currently under review, and an accompanying industrial benchmark dataset integrated as a structural chapter.