A Small Target Detection Method Based on the Improved FCN Model

Abstract

Traditional object detection is mainly aimed at large objects in images. The main function achieved is to identify the shape, color, and trajectory of the target. However, in practice, small target objects in the image must be detected in addition to large targets. Common small target detection (STD) is mainly used in intelligent transportation, video surveillance, and other fields. Small targets have small size, few pixels, and low resolution and are easily blocked. Small object detection has emerged as a research problem and a hotspot in the field of object detection. This study proposes an improved FCN model based on the full convolutional neural network (FCN) and applies it to the STD. The following is the central concept of the proposed method. Small targets are prone to occlusion and deformation in the image data. The deformation here is mainly reflected in the larger shape obtained by shooting at different angles. Therefore, it is a challenge to fully and accurately obtain the characteristics of small targets. The traditional method based on multilayer feature fusion cannot achieve ideal results for STD. This study is based on FCN and introduces a spatial transformation network. The network can optimize the handling of problems such as partial occlusion or deformation of small objects. The use of a spatial transformation network can alleviate the problem of poor feature extraction caused by partial occlusion or deformation of small targets, improving final detection accuracy. The experimental results on public datasets show that the proposed method outperforms other deep learning algorithms (DLA) in the detection accuracy of small target objects. Furthermore, the model’s training time is reduced. This study’s research provides a good starting point for the detection, recognition, and tracking of some small objects.