Visual agents operating in complex environments demand strong generalization capabilities, a key research focus in RL. However, conventional RL models typically depend on large amounts of interaction data and often struggle to adapt to semantic variations and scene shifts in unfamiliar environments. To address this limitation, this paper introduces a visual agent architecture that leverages lightweight convolutional perception and policy transfer. Specifically, the approach adopts a MobileNet image feature extractor, freezes the perception parameters, and fine-tunes the policy network to enable rapid adaptation and generalized learning across diverse scenarios. Experimental results show that the proposed method maintains about 86% of policy performance under image perturbations such as background replacement and color variations, while end-to-end CNN-PPO and ResNet18-PPO suffer performance drops of 33.3% and 36.7%, respectively. Moreover, it reduces training time by over 62%, cuts model parameters by 79.5%, and doubles inference speed, highlighting its effectiveness, efficiency, and scalability.