A Study on the Effectiveness of Detection of Unbalanced Datasets Based on Faster R-CNN
Delong Cai, Zhaoyun Zhang, Guanfeng He · 2022
In the Faster R-CNN target detection task, to address the problem of how the unbalanced distribution of different kinds of datasets affects the recognition effect of the model, a research scheme is proposed to train the model using two different ratios of dataset combination, the two ratios are 1:1 and 1:2 respectively.In this paper, two kinds of objects in electrical equipment, anti-vibration hammer and insulator, are taken as the research objects, and each type of dataset After data processing, the data sets are combined according to the two different ratios set, and then put into the model for training to get the recognition effect. The experimental results show that at Epoch=40,000, the AP(the explanation is shown in TABLE I) of the trained model is much higher for the categories with a small proportion of data than for the categories with a large proportion of data when the ratio of the data sets of anti-vibration hammer and insulator is 1:2 or 2:1, and the AP of the model trained by unbalanced data distribution is unstable compared with the AP of the model trained by the ratio of the two categories of data sets of 1:1. In the actual detection task, it is easy to cause "false detection" or "missed detection". Based on the data unbalance problem, we propose data processing methods such as undersampling and oversampling of the dataset.