Because rice pests are usually small in size， some different types of pests have similar appearance， and the same type of pests have different appearance in different growth processes， it is very difficult to identify rice pest types. We improved YOLOv7 neural network by introducing convolutional block attention module and feature pyramid module and constructed a challenging rice pest dataset， which is collected from rice planting base of Ezhou City in Hubei province to recognize rice pests. According to the characteristics of sample distribution， data enhancement was carried out， and random noise， Mixup， Cutout and other data enhancement methods were introduced to make the deep learning model learn the visual features of pest discrimination from a deeper dimension. Taking MobileNetv3 as the backbone network， the YOLOv7 network was improved， and a multi-scale neural network model based on feature pyramid was constructed to improve the identification accuracy of small individual pests. The results showed that the average accuracy rate of rice pest detection based on the improved method is 85.46%， surpassing the networks such as YOLOv7 and Efficient Net-B0. The size of the improved YOLOv7 model is 20.6 M， and the detection speed is 92.2 frames/s， which is more than 5 times that of the original YOLOv7 algorithm. The results indicate that this method can be applied for remote automatic recognition of rice pests.