An Optimization Object Detection Model on Brassica Napus Area Based on Image Compression Framework

Abstract

Using UAV (Unmanned Aerial Vehicle) equipment, it is often easy to take aerial images of brassica napus in the field, and automatically divide the brassica napus areas in the images by the trained object detection network model, which are used for the subsequent research of brassica napus flowering identification. However, the original aerial images obtained from the UAV equipment have a high resolution of about 5427×3078, and each image also takes up more memory space of about 10 MB. In the limit of hardware resource environment, especially in the case of insufficient GPU video memory, if all the original images are used to train the brassica napus object detection model, it will cost a lot of time, and the training process may also fail. To solve the above problems, a modified image compression framework based on deep learning is proposed to process the original aerial images of brassica napus in this paper and compress the storage capacity of the images on the condition of constant image resolution, so as to speed up the training process of brassica napus object detection model. After experimental analysis, the compression ratio of each original image reaches 6.34, and the training time of the brassica napus object detection model is also reduced to 58.7%, achieving the goal of reducing the training time of the model. Finally, the mAP (mean average precision) of the object detection model reaches 97.13%.