Enhancing Dog Detection in Challenging Conditions

Imagine you've engineered a robot designed to locate lost dogs. Upon spotting a canine, the robot verifies if its breed matches any recently reported missing breeds. As a proficient data scientist, you've equipped the robot with a convolutional neural network to distinguish between various dog breeds through image and video analysis. However, a challenge arises: the robot struggles to differentiate between pugs and pit bulls due to a training dataset error, where these breeds were mislabeled 50% of the time. This issue is exacerbated under suboptimal viewing conditions such as rain, fog, distance, or when the dog is partially obscured. What strategies would you employ to refine your neural network to effectively address these challenges?