Enhancing AI Systems through Representative Dataset, Transfer Learning, and Embedded Vision

Abstract

<p>Artificial intelligence (AI) encompasses a range of techniques that enable machines to perceive, learn, and make intelligent decisions and it has emerged as transformative technology in many applications. This thesis presents the development of an AI model, focusing on the significance of the primary representative dataset and the effectiveness of transfer learning and fine-tuning techniques for model development. The research demonstrates the affirmative impact of methodical approaches on the accuracy, efficiency, and robustness of AI systems. Moreover, the application of the detection model is demonstrated in wastewater management i.e., for urban wastewater systems, thus underpinning the application of AI to real world scenarios. <p>The research approach followed in this work includes critical literature review, site surveys, intensive experimentations, and robust validation processes which allowed to identify and address existing gaps and limitations and helped to develop AI detection models for the selected application. <p>Deep neural networks, a prominent AI technique, chosen for developing AI model in this work has exceptional capabilities in handling complex tasks by learning from vast amounts of data. But the availability of high-quality and representative datasets to effectively train deep neural network models is critical. The comprehensive and diverse datasets provide effective training examples, reduce biases, and enhance the detection models’ ability to handle complex inputs. <p>In the present case, the representative dataset was not available. Therefore, critical multiclass representative image dataset was generated in the laboratory with unparalleled authenticity using model sewer network and named as Sewer-Blockages Imagery Recognition Dataset (S-BIRD) which served as a benchmark for real-time detection and recognition models. The research also addressed the need for dataset curation, data integrity, and biases. <p>Using S-BIRD, deep neural object detection models were developed through transfer learning and fine-tuning. Inductive transfer learning technique used for development of models, improved convergence, training times, and performance on target detection tasks, enabling adaptation to different domains with minimal additional training. Transfer learning parameters were optimised for desired outcomes. The effectiveness of the developed model for detecting sewer blockages was evaluated by performance metrics. The model achieved high accuracy rate of 96.30% at an IoU of 0.5 in detecting different blockages validating efficacy of dataset and the applicability of the techniques used for developing the model. <p>AI detector trained on the S-BIRD dataset was then imported on advanced GPU-based single-board computer that formed an embedded vision-based automation system for the detecting sewer blockages. The output of the present research contributes to the advancement of AI and its application in wastewater management. The knowledge and findings acquired from this research form a strong foundation for future explorations and advancements in the AI field and facilitating its widespread implementation across various domains. <p>For future research work integration of AI techniques like semantic segmentation, instance segmentation and panoptic segmentation, can be investigated to reinforce detection tasks. To enhance model robustness, expansion of representative datasets coupled with continuous learning approaches is recommended. For further practical application of the outcome of the thesis, collaboration with industry will yield advancements in AI innovation.