Accurately detecting student classroom behaviors in classroom videos is beneficial for analyzing students’ classroom performance and consequently enhancing teaching effectiveness. To address challenges such as object density, occlusion, and multi-scale scenarios in classroom video images, this paper introduces an improved YOLOv8 classroom detection model. Firstly, by combining modules from the Res2Net and YOLOv8 network models, a novel C2f_Res2block module is proposed. This module, along with MHSA and EMA, is integrated into the YOLOv8 model. Experimental results on a classroom detection dataset demonstrate that the improved model in this paper exhibits better detection performance compared to the original YOLOv8, with an average precision (mAP@0.5) increase of 4.2%.

With the advancement of technology, the demand for increased production efficiency has gradually risen, leading to the emergence of new trends in agricultural automation and intelligence. Precision classification models play a crucial role in helping farmers accurately identify, classify, and process various agricultural products, thereby enhancing production efficiency and maximizing the economic value of agricultural products. The current MobileNetV2 network model is capable of performing the aforementioned tasks. However, it tends to exhibit recognition biases when identifying different subcategories within agricultural product varieties. To address this challenge, this paper introduces an improved MobileNetV2 convolutional neural network model. Firstly, inspired by the Inception module in GoogLeNet, we combine the improved Inception module with the original residual module, innovatively proposing a new Res-Inception module. Additionally, to further enhance the model's accuracy in detection tasks, we introduce an efficient multi-scale cross-space learning module (EMA) and embed it into the backbone structure of the network. Experimental results on the Fruit-360 dataset demonstrate that the improved MobileNetV2 outperforms the original MobileNetV2 in agricultural product classification tasks, with an accuracy increase of 1.86%.