Image recognition of pear leaf diseases is an important task of plant protection. The lesion area of pear leaf diseases is not fixed in the whole leaf, which has the characteristics of randomness. The convolution neural network is used to identify the images of pear leaf diseases, due to its rotation and translation invariance, the generalization ability of the model is weak. The capsule network uses feature vectors to replace feature value, and uses dynamic routing to replace pooling to obtain spatial information between entities. However, the size of lesion area of pear leaf diseases is random, and the capsule network cannot fully extract features, resulting in a decrease in recognition rate. To solve the problem, a pear leaf diseases image recognition model based on capsule network was proposed, which uses conditional convolution to customize specific convolution kernels for each input to adapt to pear leaf diseases images of different sizes. The experimental results show that the recognition accuracy, precision, recall and F1score of the proposed algorithm are 91.33%, 91.40%, 91.33% and 91.36%, which are better than capsule network.