Abstract: Apples and tomatoes are very common fruits and vegetables in our daily life. Accurate identification of diseases can improve crop yields and reduce economic losses. Aiming at the problem that existing plant disease detection methods cannot accurately and quickly detect diseased areas in plant leaves, a deep learning method based on improved Yolov5 is designed to detect common diseases of apple and tomato leaves. The data set of apple and tomato leaf disease was constructed by data enhancement and image annotation technology, and the k-means algorithm was used to adjust the initial anchor frame. On this basis, the composite backbone network was used to enhance the disease feature extraction ability of Yolov5 backbone network, and the Varifocal Loss function was used to improve the identification accuracy of densely infected areas. The test results show that the mAP of the improved Yolov5 disease detection algorithm reaches 95.7%, and the mAP is increased by 1.7% on the basis of the original Yolov5 model. The average detection time of an image is 0.033 s, which provides a high-performance method for apple and tomato leaf disease detection. The solution can classify and locate plant leaf diseases with high accuracy.