Abstract: With the continuous development of recycling technology for urban mineral resources, recycling of used mobile phones has become a hotspot for current research. Restricted by the relative lack of computing resources and data resources, the accuracy of used mobile phone recognition based on current off-line intelligent recycling equipment is difficult to meet the practical application. Therefore, an image identification method based on multivariate feature heterogeneous ensemble deep learning method was proposed. First, the character region on the back of the mobile phone was extracted by using character region awareness for text detection (CRAFT) algorithm, the VGG19 model pre-trained by ImageNet was used as the image feature embedding model, and the local character feature and global image feature were extracted by using the transfer learning mechanism. Then, the optical character recognition (OCR) character recognition model based on neural network (NN) mode was constructed by using the local feature, and the improved deep forest classification (DFC) model based on non-NN model was constructed by using the global and local features. Finally, the outputs of heterogeneous OCR and the DFC model were integrated and fed into the Softmax to ensemble, and the final recognition result was obtained based on the criterion of maximum category weight vector. The effectiveness of the proposed method was verified based on real images of used mobile phone from recycling equipment.