An Optimization Scheme of Object Detection Model Based on CNN Feature Visualization Method

In the field of computer vision, enhancing the interpretability of networks is a hot research issue. We proposed a feature visualization interface based on Grad-CAM algorithm, which can match the convolutional neural network under Pytorch framework and TensorFlow framework. We added the SE attention module and SPP module and analyzed the phenomenon of object detection by using the feature visualization interface. At the same time, considering the deployment on resource-limited equipment, 8.9M and 109M YOLOv3s and YOLOv3m models are designed, and the phenomenon of object detection is analyzed by using the feature visualization interface. When the model size of YOLOv3s model is reduced by 96.2%, the accuracy of object detection can reach 0.676, which meets the deployment condition of resource-limited equipment.