The electronic equipment to perform properly, a printed circuit board's (PCB) quality is crucial. Significant research and development efforts were made to automate PCB inspection for flaw detection, mostly using computer vision techniques, in order to meet the necessary quality standards. Despite these developments, different board and component sizes frequently affect how accurate these procedures are. Attempts to improve its accuracy, particularly for minute or minor flaws on a PCB, frequently result in a trade-off of decreased real-time performance, which restricts its use in the manufacturing sector. In order to overcome the challenge of real-time inferring small or variable flaws on a PCB, this research proposes an improved deep learning network. We propose two advancements: (i) a refined complete intersection over union loss function to accurately capture small flaws and (ii) a novel multi-scale feature pyramid network to improve tiny defect identification by including context information. Our model maintains real-time inferencing performance at 90 frames per second while achieving 99.17% mean-average precision, according to experimental results on a publicly available dataset of PCB faults.