An adaptive visual diagnostic model for complex scenes based on unsupervised pseudo-label continuous learning and BiFPN-enhanced YOLO.

With the deployment of intelligent systems in open environments, visual perception capabilities are required for engineering applications. Traditional static visual detection models experience feature degradation under environmental changes. Conventional pseudo-labeling techniques exhibit confirmation bias when processing complex samples, and models encounter knowledge forgetting during the continuous absorption of unlabeled data streams. This paper constructs an adaptive visual diagnostic framework for complex scenes, integrating unsupervised domain adaptation with continuous learning. At the feature extraction level, by introducing a BiFPN node, the network compensates for multi-scale spatial distortions caused by environmental changes while maintaining computational economy. A pseudo-label bidirectional recovery mechanism constrained by dynamic thresholds is utilized to extract supervision signals and alleviate confirmation bias. During the parameter iteration phase, a structured predictive distillation loss function is integrated to mitigate the knowledge forgetting phenomenon during continuous data absorption. Tests on the BDD100K dataset show that, with a computational load of 16.6 GFLOPs and an inference speed of 19.3 FPS, the model records an mAP@0.5 of 43.5% and 47.1%, and an mAP@0.5:0.95 of 22.1% and 24.5% in the illumination and meteorological domain transfer tasks, respectively. In the Cityscapes fog transfer test, the model reaches an mAP@0.5 of 48.3% and an mAP@0.5:0.95 of 25.1%. After 10 rounds of incremental evolution, the source domain mAP@0.5 remains at 52.8%, demonstrating knowledge retention capabilities comparable to baseline continuous learning methods.