Fluorescence imaging has demonstrated great potential for malignant tissue inspection. However, poor imaging quality of medical fluorescent images inevitably brings challenges to disease diagnosis. Though improvement of image quality can be achieved by translating the images from low-quality domain to high-quality domain, fewer scholars have studied the spectrum translation and the prevalent cycle-consistent generative adversarial network (CycleGAN) is powerless to grasp local and semantic details, leading to produce unsatisfactory translated images. To enhance the visual quality by shifting spectrum and alleviate the under-constraint problem of CycleGAN, this study presents the design and construction of the perception-enhanced spectrum shift GAN (PSSGAN). Besides, by introducing the constraint of perceptual module and relativistic patch, the model learns effective biological structure details of image translation. Moreover, the interpolation technique is innovatively employed to validate that PSSGAN can vividly show the enhancement process and handle the perception-fidelity trade-off dilemma of fluorescent images. A novel no reference quantitative analysis strategy is presented for medical images. On the open data and collected sets, PSSGAN provided 15.32%~35.19% improvement in structural similarity and 21.55%~27.29% improvement in perceptual quality over the leading method CycleGAN. Extensive experimental results indicated that our PSSGAN achieved superior performance and exhibited vital clinical significance.