Research on Network Migration and Performance Optimization of the YOLOv7 Model Improved Based on MindSpore Framework

As an important branch in the field of computer vision, object detection plays an increasingly important role in fields such as autonomous driving, security monitoring, and medical image analysis. YOLOv7, as one of the current advanced object detection models, has received extensive attention due to its characteristics of high efficiency and accuracy. However, when YOLOv7 is adapted to the MindSpore framework, there are performance bottlenecks. Aiming at the real-time object detection requirements of edge devices, this paper proposes an improved YOLOv7 model based on the MindSpore framework. By staticizing the shape of dynamic tensors, reconstructing the backbone network with Depthwise Separable Convolution, and fusing the hybrid attention mechanism, the inference efficiency of the model on the Ascend chip is optimized. Experiments show that the [email protected] of the improved model on the COCO dataset is increased by 3.2%, the inference speed reaches 142 FPS (an increase of 23% compared with the original model), and the number of parameters is reduced by 28%. This research contributes to the development of efficient object detection within the MindSpore ecosystem.