A progressive sampling method for object detection performance surface based on Gaussian process multi-kernel fusion

With increased interest in robustness evaluation of deep learning models, performance assessments under single-dimensional perturbations have been extensively studied, resulting in the establishment of numerous benchmarks. However, the behavior of models under bi-dimensional perturbations remains underexplored. A key issue arises from the exponential growth in sampling requirements when modeling performance surfaces in two-dimensional perturbation spaces, resulting in significant computational overhead. To address this issue, we propose a progressive sampling method for object detection performance surfaces that uses multi-kernel Gaussian process fusion. Our method incorporates a genetic algorithm to optimize kernel composition, leveraging the superior surface fitting capabilities and uncertainty quantification of composite kernel Gaussian processes. A reinforcement learning strategy is used to generate an initial population with high diversity and broad coverage. In addition, a whale optimization algorithm is used to fine-tune the weights and parameters of individual kernels, thereby improving sampling efficiency. Experimental results show that the proposed method significantly improves the sampling efficiency of performance surfaces, effectively reducing the number of samples required. This provides a reliable and efficient solution for robustness evaluation of deep learning models under complex perturbation scenarios.