In multi-scenario perception and intelligent control systems, the accurate and efficient recognition of traffic light states is an important support for ensuring stable system operation and the collaborative operation of intelligent devices. In response to the diverse appearances and scarce samples in traffic light detection scenarios, transfer learning has become a mainstream solution. However, traditional full parameter fine-tuning and shallow freezing do not quantify the specific contributions of network layers to the task, leading to overly broad training updates, parameter redundancy, and a tendency to overfit under small sample conditions, which limits model performance. To address this, this paper proposes a layer-adaptive guided freezing strategy generation method based on CNN-LRP. By utilising the forward correlation ratio, spatial focus degree, and layer depth penalty factor, the pixel-level correlation of CNN-LRP is expanded to the convolutional layer level, constructing a layer contribution evaluation system that integrates interpretative analysis with the freezing strategy. To tackle the issue of insufficient task adaptability in traditional strategies, this paper designs a multi-metric fusion convolutional layer adaptability scoring method that automatically generates freezing strategies, and realises task-driven migration by freezing the low-adaptability layer, so as to optimise the model. Finally, verification tests were conducted using scene data with different lighting and distances. The results show that compared with full fine-tuning and other freezing strategies, the model generated by the method proposed in this paper has an improvement of 3.1% in test accuracy, a reduction of 34.49% in parameter tuning volume, and better performance in terms of accuracy, efficiency, generalization, robustness, and the universality of convolutional networks.