A NOVEL APPROACH TO OBJECT DETECTION IN REMOTE-SENSING IMAGES BASED ON YOLOv3

Hyperspectral images can obtain an approximately continuous spectral curve of a ground object in ultraviolet, visible, near infrared, and mid infrared. Through the target detection of hyperspectral images, the important ground objects can be obtained. In practical applications, based on the poor intelligence and robustness of traditional high-resolution, remote-sensing image target detection algorithms, and compared with ResNet-152 and ResNet-101, the YOLOv3 algorithm is improved to obtain better feature extraction capabilities and powerful detection performance. This paper applies extended convolution, label smoothing, and an improved non-maximum suppression algorithm (NMS) to soften the non-maximum suppression algorithm (soft NMS). The experimental results show that the average accuracy of the improved YOLOv3 is 95.72%, which is higher than the mainstream target detection networks such as fast-RCNN and faster-RCNN. Moreover, the average accuracy is 5.39% higher than that of the traditional YOLOv3. The improved YOLOv3 network shows higher average accuracy in the detection of small and overlapping targets.