LightYOLO-Touch: A Lightweight Object Detection Framework for Visual and Tactile Image Analysis

Lightweight object detection models are essential for practical perception tasks in resource-constrained environments such as embedded robotics and tactile sensing. In this paper, we present LightYOLO-Touch, a compact and efficient detection framework tailored for both visual and tactile image analysis. Based on an improved YOLOv8n architecture, the model replaces standard convolutions with GhostConv to reduce redundancy, and introduces a C2fCIB module for enhanced multi-scale feature extraction. To improve spatial sensitivity—especially critical in low-texture tactile inputs—we incorporate Dual Coordinate Attention Feature Extraction (DCAFE) modules to strengthen directional and positional awareness. A lightweight shared detection head, Detect_LSCD, enables efficient multi-resolution prediction with fewer parameters. Additionally, we apply a modified Wise-IoU loss (Wise-IoU v3) to improve bounding box regression by balancing localization accuracy and training stability. Extensive experiments on the COCO2017 dataset and a custom tactile dataset demonstrate competitive accuracy with reduced computation, and real-time deployment on a Raspberry Pi 4B confirms its practical viability in embedded environments.