Lightweight object detection algorithm based on improved FCOS

Abstract

Fully Convolutional One-Stage Object Detection (FCOS) is a one-stage anchor-free object detection model. The detection accuracy even exceeds some two-stage and anchor-base object detection model, but it still has the problem of slow inference speed. Aiming at the problem that the speed of FCOS algorithm cannot meet the requirements of real-time detection, a lightweight object detection model Improved MobileNetV2-FCOS (IM-FCOS) is proposed. First of all, the MobileNetV2 structure is introduced as the backbone network of FCOS. By replacing the ordinary convolution with a depth separable convolution, the model size is reduced while improving the model detection speed. Then, use the Sobel module to add the directional gradient extracted by the Sobel convolution, increase the dimension of the input data, and enhance the edge semantic information. Using the Mish activation function to replace the Rectified Linear Unit (ReLU) activation function in the Head part of the model, retain the semantic information of the negative part, and improve the model generalization ability. Use Distance Intersection over Union (DIoU) as an evaluation indicator to more accurately describe the similarity between the predicted frame and the ground truth, and optimize the regression loss. Optimize the model algorithm from the above three aspects, and finally the experimental results on the Microsoft Common Objects in Context 2014 (MS-COCO2014) data set show that the mean Average Precision (mAP) of the IM-FCOS algorithm reaches 34.1, and Frames Per Second (FPS) reaches 31, which can meet the need for real-time detection.