Visual perception is the primary source for humans to acquire information. The mimicking of visual systems is crucial to develop artificial intelligence technologies. Currently, optoelectronic synapses are widely used in artificial visual systems due to the in-memory processing of optical signals. However, the photoelectric conversion of the synapses requires contact processing of input optical signals, which leads to significant energy consumption. In this paper, all-optical artificial synapses were presented based on photochromic perovskite thin films. Under UV and visible light pulse stimulation, the perovskite films exhibit synaptic behaviors in optical transmittance changes, including paired-pulse facilitation and learning ability. Through a recurrent neural network processing the time-dependent transmittance change data, a 100% accuracy in the classification of two digital images can be instantly achieved, even in the first epoch. The all-optical synapses provide an innovative pathway toward energy-friendly artificial visual systems.