Open lq132069 opened 11 months ago
Dear friend,
I trust this message finds you in good health. I'd like to express my sincere gratitude for your question and keen interest in our work.
The issue of imbalance of categories does exist in the field of hyperspectral anomaly detection, e.g., anomalies are in a low percentage of the images. Our model is trained in a self-supervised learning manner, which does not require any a priori information about the background and anomalies, and only utilizes unlabeled original images, with the expectation that the model can become a good background reconstructor, i.e., the model will reconstruct the background well but not anomalies, resulting in a large reconstruction error for anomalies. In the field of hyperspectral anomaly detection, the fact that the number of background pixels is much higher than that of anomalous pixels contributes to the trend that the network tends to reconstruct the background well, and hence this is beneficial for performing hyperspectral anomaly detection.
Thank you for your thoughtful engagement. Best regards, Degang Wang
Dear Mr. Wang,
Hope this issue find you well.
May I ask for a issue for the training process? How to solve the imbalance of categories in the training data set? I'm failed. the positive class and negative class is extremely imbalance, like 1:80.
Thank you for your time. Warm Regards, your fans.