In this paper, deep learning is used to study image recognition techniques. Firstly, the image recognition process is structured, the YOLOv4 network framework is constructed, the features are extracted using the PANet reinforcement network, and the image overlap is extracted using the loss function. Then, we make an improved architecture ACDNet algorithm based on YOLOv4 and set the main function of the ACDNet model. Finally, the accuracy of image recognition under different algorithms and the recognition effect evaluation of the ACDNet algorithm are tested, respectively. The study shows that the image recognition accuracy of the ACDNet algorithm is located in the first of the three algorithms, with the highest accuracy of 98.16%, which is good and effective for image recognition and classification. The accuracy of ACDNet in the training set of plant image recognition is 99.34%, which is good for classification and recognition performance.