The paper addresses the un-explored scenario in intelligent agriculture and computer science, e.g., pesticide label detection. The problem opens to an exciting challenge in image recognition where the deployed system heavily depends on the performance of machine learning models despite unconstrained environments. To build up the system, the authors collect a real-world dataset to evaluate several state-of-the-art object detection algorithms. The authors select a dataset of 1221 photos containing 32 common pesticides on mango trees. Then we evaluate off-the-shelf deep convolutional networks to detect pesticide labels and take into account the detection accuracy. Finally, we integrate the best model into our self-developed mobile application that (i) correctly detects pesticide labels online and offline and (ii) provides essential pesticide information to facilitate further integrated treatment and services.