With the progress and development of science in today's era, fungi have become an indispensable part of the national diet structure. However, because of this, different fungi can easily be mistaken for food, leading to the occurrence of some diseases. However, in the production environment, the classification and identification of edible fungi is still mainly based on manual identification and classification. Manual identification relies too much on experience, and requires a certain cost in manual learning and training. Moreover, manual identification has low efficiency, waste of manpower and a certain error rate, which leads to the increase of certain costs in the production and sales of mushrooms. Therefore, the target detection classification recognition is applied to the mushroom recognition classification, the use of efficient and advanced automated recognition process to replace the traditional manual recognition, help to make the mushroom classification in the production and sales link more accurate, save time, save labor consumption and money and other costs. However, the current deep learning target detection is applied to the study of mushroom recognition, there are few types of recognition, and there are many types of mushrooms in the recognition screen can not accurately locate the target classification. This paper proposes an improved yolov5 algorithm to identify mushrooms (MR-yolo), using different equipment and ways to collect 21 types of edible mushroom images and make data sets, in the case of maintaining the network depth and width of the original yolov5 algorithm network structure is modified to achieve the improvement of the performance of the target detection algorithm.