A Training Strategy to Optimize the Performance of Target Detection Model
Main Article Content
Abstract
Although neural networks have made tremendous progress in feature extraction, each year new models refresh the accuracy of previous models on major data sets, which also shows that there are still a considerable number of effective features in the data set. This paper presents a model and a training strategy to fully exploit the effective features in the data set. The main feature of the model is the use of focal loss based on gradient cropping; the main steps of the strategy include: (1) build a model that contains all necessary optimization techniques, and debug the model capacity to the highest accuracy, get an initial model; (2) corrects the missing and wrong labels of the data set based on the initial model to obtain the training set 1, training the model, and obtaining the model 1; (3) extracting the FN and FP as the training set 2, training the model, and obtaining the model 2; (4) repeating (3) ; (5) NMS is used to merge the prediction results of filtering model 1 to model n. Experiments show that, based on the model obtained by this strategy, MAP increases by 2.7%.