Abstract: Balsam pear is a type of vegetable that is popular with humans, which is rich in crude fiber, vitamin C, calcium, iron, and other nutrients, therefore, it is highly significant for human health, and the cultivated area of balsam pear increases annually, however, it is more susceptible to diseases that affect the quantity and quality of the crop yield. Recently, it was reported that the Balsam pear is exposed to many diseases namely, powdery mildew, gray leaf spot, gummy stem blight, Phyllosticta leaf spot, etc. The aim of the study was an automatic identification method for Balsam pear leaf diseases based on improved Faster R-CNN. 1 204 pictures of balsam pear leaves were photographed under sunlight which included the healthy balsam pear leaf as control and infected leaves with powdery mildew; gray leaf spot; gummy stem blight, and Phyllosticta leaf spot. The data were accumulation and rotating and flipping randomly. Finally, 10 627 pictures were obtained as the experimental groups. The design of the experiment was carried out as 9 564 pictures used as a training group, 1 435 pictures of the training group were used as a detection group. and 1 063 pictures were used as the experimental groups. The residual-structure convolutional neural network, ResNet-50, was used as the feature extraction network of Faster R-CNN. Two classic networks, ZF-Net and VGG-16, were compared to extract features that showed significant results, proving whether the proposed approach performed well or not. The pre-trained ImageNet model for transfer learning was joined to the network, which saved computing and time costs. Feature maps obtained from the feature extraction network were input to the Region proposal network for proposals, which included a series of anchors, scores, classification loss, and bounding-box regression loss. Finally, feature maps and proposals were sent to R-CNN for the ultimate accurate location. The obtained results indicated that the performance of the original model was not good because balsam pear leaf diseases were complicated in color, texture, and shape. The feature extraction networks were likely to miss some diseases of the balsam pear leaf diseases because their forms were tiny, and some took up a few pixels that made it difficult to distinguish diseases. To solve the difficulty of detecting the small targets of balsam pear leaf diseases, Faster R-CNN was modified to increase the different sizes of bounding boxes, along with the introduction of the feature pyramid networks to ResNet-50. Feature pyramid networks helped to extract feature maps from every block in ResNet-50, which contained an accurate location and strong semantic information. Feature pyramid networks then handed it over to the Region proposal network to get more accurate proposals to make object recognition. As a result, the experiment showed that the performance of the trained model using the original ResNet-50 was better than that used ZF-Net or VGG-16 as a feature extraction network, and the mean average precision was 78.85%. The average precision of healthy balsam pear leaves, powdery mildew, gray leaf spot, gummy stem blight, and Phyllosticta leaf spot were 88.24%, 75.73%, 51.75%, 72.00% and 72.15%, respectively. The mean average precision for the feature pyramid networks’ final model was 0.863 9, which is 7.54% higher than before. The average precision of healthy balsam pear leaves, powdery mildew, gray leaf spot, gummy stem blight, and Phyllosticta leaf spot were 89.24%, 81.48%, 83.31%, 88.62%, and 89.24%, respectively, and the average precision of gray leaf spot increased the stain by more than 16%. Detection time reached 0.322 s/frame, ensuring real-time detection. This method was characterized by good durability and high accuracy to detect diseases of balsam pear leaf diseases in a complex natural environment and had important research significance for the prevention of melon and fruit diseases.