Enhancing Common Loss Functions: “Append” with Overall Metrics as Supplementary

In image processing, the well-designed loss functions can effectively guide the model towards training goals, which play a crucial role in defining the optimization objectives of neural network training. However, commonly used loss functions are widely applied among them due to their simplicity and effectiveness. Although many literatures have discussed the limitations of commonly used loss functions, which consider few features of information for their relatively simple design paradigm, and proposed more effective alternatives, the commonly used loss functions still remain popular due to their simplicity in conceptualization and application. Hence, devising strategies to enhance the performance of the common loss functions while keeping their simplicity and practicality has became a prominent research focus. This paper proposes a guiding principle called “Append” inspired by the methods in the field of image quality assessment that employs overall metrics to design loss functions as complements for commonly used ones. And the approaches based on the guiding principle will be collectively referred to as Append Loss. We also provide a generic append loss as an example belongs to Append loss. The experimental results in this paper demonstrate that by combining the commonly used loss functions and append loss, faster convergence speed and better results can be achieved. Due to the simplicity of the “Append”, the loss based on it will remain simple and easily generalizable.