This thesis presents a temperature-dependent reference control method to compensate for the temperature effect on CMOS image sensors from -40◦C to 125◦C. Recently, machine vision has been one of the most important applications for CMOS image sensors. However, the working environment and operation might generate large temperature variations and degrades the performance of CMOS image sensors. In this work, the temperature dependency of the distortion generated in the analog front-end is investigated. A mathematical model has been built to describe the relationship to its temperature dependency. Besides, a temperaturedependent reference control method with 16 different slopes is proposed. This method can control the working condition of the transistors and compensate for the change by reference current or voltage. Besides, the slope design can cover the external noise or mismatch and fulfill the compensation. Finally, it can compensate the signal distortion and prevent the settling error which could generate FPN while maintaining the noise performance. A SPICE simulation and post-simulation are performed to confirm the results of the design.

Neonatal sepsis is a dangerous non-specific disease in babies, especially neonate/newborns. It is one of the leading causes of neonate mortality rate, because of the difficulty to diagnose, leading to late or false treatment. Previous research has found the promising feature of artificial intelligence or machine learning in solving the problem. After analysing hours of the electronic health record data available, they are able to diagnose sepsis condition on neonates. However, the accuracy and time needed before diagnosis are still concerning considering the risk of mistreated or late diagnosed sepsis cases. In this research, machine learning and thermal imaging technology is used to explore the possibility of predicting sepsis. 57 thermal videos from 26 babies are processed to track the highest skin temperature visible to the thermal camera. The temperature data then is utilized to train and test several machine learning models for predicting sepsis cases. Support Vector Machine (SVM) was found to be the best sepsis predictor using time-series variation of the temperature data as the feature. The model needs 10-30 minutes of thermal recording, 19 minutes in average, to predict sepsis and achieved 82\% accuracy. Simulation also shows the high possibility in increasing the accuracy when more data/thermal videos are available to train the model. High accuracy model with fast reacting sepsis prediction could help doctors precisely treat septic neonates in timely manner, decreasing the mortality rate for sepsis cases.