【摘要】 找出目标的位置和类别是目标检测的主要任务。随着人工智能和深度学习的发展，目标检测可以达到人眼所达不到的精度。由于信息较少，覆盖面积小且基于锚框的检测算法易受锚框大小、比例数目的影响，对较小的目标难以精确检测。针对以上问题，改进无锚框算法全卷积单阶段目标检测（Fully Convolutional One-stage Object Detection,FCOS）实现了小目标检测的效率和精度。将FCOS算法的特征提取网络结构残差网络（Residual Network,ResNet）更换为轻量级网络结构MobileNetV3，随后在骨干网络中引入通道注意力机制和空间注意力机制对特征提取网络进行改进，最后设计T交并比（TIOU）代替原本的交并比（IOU），改善模型精度。实验结果表明，所改进的网络结构与FCOS相比，网络训练时间和模型大小为原来的一半，计算参数量由原来的32.12×10~6减少为11.73×10~6，减少到原来的三分之一，模型推理速度提升了10%，每秒传输帧数为11帧，与主流网络FasterRCNN相比，检测精度和速度更快，可以满足对小目标的实时检测。更多还原

【Abstract】 The main task of target detection is to find the location and category of the target. With the development of artificial intelligence and deep learning, target detection can achieve accuracy beyond human eyes. Due to less information and small coverage area, the detection algorithm based on anchor frame is easily affected by the size and proportion of anchor frame, and it is difficult to accurately detect small targets.Aiming at the above problems, Fully Convolutional One-stage Object Detection(Fully Convolutional One-stage Object Detection, FCOS) algorithm without anchor frame is improved to achieve the efficiency and accuracy of small target detection. FCOS feature extraction network structure ResNet(Residual Network,ResNet) is replaced with lightweight network structure MobileNetV3. Then, channel attention mechanism and spatial attention mechanism are introduced in BackBone network to improve feature extraction network.Finally, T-intersection ratio(TIOU) is designed to replace the original intersection ratio(IOU) to improve model accuracy. Experimental results show that compared with FCOS, the network training time and model size of the improved network structure are half of the original, the number of calculation parameters is reduced from 32.12×10~6 to 11.73×10~6 which is one third of the original, the model inference speed is increased by 10%, and the transmission frame per second(FPS) is 11. Compared with the mainstream Faster network Faster RCNN detection accuracy and speed, can meet the real-time detection of small targets.更多还原