An Improved Image Classification Method for Specific Reality Environments

This paper proposes an improved image classification method based on specific data features and practical product requirements in the real enterprise environment. In order to make the model capable of filtering images outside the predefined classes and increasing the precision of the model, two probability thresholds are set at the last layer of the model, and "the uncertain class" is introduced to control the output of the model. The predefined classes are subdivided to sub-classes in training procedure and reassemble after training, which improves the stability and robustness of the model. A parallel architecture is also proposed to improve the efficiency of the entire system. The model is evaluated on a real image dataset collected by ourselves. Compared with the traditional deep model, the proposed method achieves faster convergence, higher accuracy and stability. Especially, it can filter out most of uncertain classes that traditional models cannot recognize. The system is already available in the practical enterprise environment, and it is highly efficient and extendable.