SSIT: a sample selection-based incremental model training method for image recognition

In the big data environment, the expansion of image data sets makes the image recognition process need to adapt to sample characteristics and data distribution changes. In this case, image recognition research focuses on finding the balance point of incremental learning in the stability-plasticity dilemma under limited computing and storage resources. The existing incremental learning methods have disadvantages in generalization performance, iteration rounds, convergence speed, and data category imbalance, so it is essential to study the incremental learning methods for image recognition training. In this study, a sample selection-based incremental model training method is proposed for image recognition. The training process is improved by optimizing the training samples needed for each iteration. A generalization error-based category determination method is proposed to avoid the imbalance of training samples. A sample selection method based on dynamic weight is proposed to avoid the problem of increasing recognition gain. At last, experiments show that this method can enhance the generalization ability of the model. At the same time, it can meet the goal of balancing the recognition effect, reducing the number of iterations, and accelerating convergence.