计算机科学
人工智能
计算机视觉
目标检测
感受野
滤波器(信号处理)
特征(语言学)
特征提取
相似性(几何)
噪音(视频)
模式识别(心理学)
对象(语法)
领域(数学)
对比度(视觉)
方向(向量空间)
传输(电信)
干扰(通信)
人类视觉系统模型
编码(内存)
编码(集合论)
视觉对象识别的认知神经科学
面子(社会学概念)
人脸检测
背景噪声
调制(音乐)
可视化
机器视觉
视觉感受
图像分割
视觉处理
图像处理
对象类检测
分割
杂乱
作者
Xiongzhen Wang,Chuan Lin,Yongcai Pan,Ruopu Wang
标识
DOI:10.1109/tcsvt.2025.3628010
摘要
Small object detection is a significant and challenging research area in aerospace. Small objects often face issues like background interference and inter-class similarity due to their small size and low boundary contrast in complex environments. Physiological studies indicate that the visual pathway’s neuronal mechanisms can effectively extract features such as contours, shapes, and colors, filter out background noise and thus recognize complex forms. Therefore, this paper proposes a remote sensing small object detection model inspired by neuronal mechanisms in visual pathways, called RSVDet. RSVDet simulates the information transmission of the ventral visual pathway (Retina-LGN-V1-V2-V4-IT) and meticulously models the involved visual areas. First, inspired by the “Retina-LGN-V1” pathway, we designed a feature enhancement module to capture low-level information. Second, based on the global-local receptive field mechanism of V2 neurons, we developed a feature extraction module for shape information. Additionally, inspired by the self-regulation mechanisms of V4 neurons, we created a self-feedback attention module to filter background noise. Finally, drawing from the orientation selectivity of IT neurons, we designed a hierarchical modulation detection head module to extract complex shape features. RSVDet achieves an AP50 of 50.1% on the Visdrone dataset with 1.72M parameters, achieving the best performance among lightweight models.The code for RSVDet can be found at:https://github.com/wxz0426/RSVDet/tree/main.
科研通智能强力驱动
Strongly Powered by AbleSci AI