In order to avoid the influence of illumination condition, overlap and other occlusion on image recognition, an improved LeNet convolution neural network is used to improve the structure of the traditional content-based recognition method. An Apple target recognition model based on the improved LeNet convolution neural network is designed and used to avoid the influence of illumination condition, overlap and other occlusion factors on image recognition. The model trains and validates Apple images in different scenarios. The results show that the network model can effectively recognize apple images. The recognition rates of independent fruits, occluded fruits, overlapping fruits and adjacent fruits are 96.25%, 91.37%, 94.91% and 89.56% respectively, and the comprehensive recognition rate is 93.79%. Compared with other methods, this algorithm has stronger anti-jamming ability, faster image recognition speed and higher recognition rate.