Maize is one of the most important global crops. It is highly susceptible to diseases during its growth process, meaning that the timely detection and prevention of maize diseases is critically important. However, simple deep learning classification tasks do not allow for the accurate identification of multiple diseases present in a single leaf, and the existing RT-DETR (Real-Time Detection Transformer) detection methods suffer from issues such as excessive model parameters and inaccurate recognition of multi-scale features on maize leaves. The aim of this paper is to address these challenges by proposing an improved RT-DETR model. The model enhances the feature extraction capability by introducing a DAttention (Deformable Attention) module and optimizes the feature fusion process through the symmetry structure of spatial and channel in the SCConv (Spatial and Channel Reconstruction Convolution) module. In addition, the backbone network of the model is reconfigured, which effectively reduces the parameter size of the model and achieves a balanced symmetry between the model precision and the parameter count. Experimental results demonstrate that the proposed improved model achieves an mAP@0.5 of 92.0% and a detection precision of 89.2%, representing improvements of 7.3% and 8.4%, respectively, compared to the original RT-DETR model. Additionally, the model’s parameter size has been reduced by 18.9 M, leading to a substantial decrease in resource consumption during deployment and underscoring its extensive application potential.

Millimeter-wave (mmWave) radar has become an important research direction in the field of object detection because of its characteristics of all-time, low cost, strong privacy and not affected by harsh weather conditions. Therefore, the research on millimeter wave radar object detection is of great practical significance for applications in the field of intelligent security and transportation. However, in the multi-target detection scene, millimeter wave radar still faces some problems, such as unable to effectively distinguish multiple objects and poor performance of detection algorithm. Focusing on the above problems, a new target detection and classification framework of S2DB-mmWave YOLOv8n, based on deep learning, is proposed to realize more accuracy. There are three main improvements. First, a novel backbone network was designed by incorporating new convolutional layers and the Simplified Spatial Pyramid Pooling - Fast (SimSPPF) module to strengthen feature extraction. Second, a dynamic up-sampling technique was introduced to improve the model’s ability to recover fine details. Finally, a bidirectional feature pyramid network (BiFPN) was integrated to optimize feature fusion, leveraging a bidirectional information transfer mechanism and an adaptive feature selection strategy. A publicly available 5-class object mmWave radar heatmap dataset, including 2,500 annotated images, were selected for data modeling and method evaluation. The results show that the mean average precision (mAP), precision and recall of the S2DB-mmWave YOLOv8n model were 93.1% mAP@0.5, 55.8% mAP@0.5:0.95, 89.4% and 90.6%, respectively, which is 3.3, 1.6, 4.5 and 7.7 percentage points higher than the baseline YOLOv8n network without increasing the parameter count.