<p>This paper presents a vision-based tactile sensor capable of estimating both contact force magnitude and direction by capturing grayscale variation and marker displacement caused by elastomer deformation. A camera records real-time images under white light illumination. An improved ResNet50-based dual-branch network is designed, with one branch for force magnitude regression and the other for direction classification, sharing low-level features for greater efficiency. Experiments show that the model achieves an average magnitude prediction error below 0.12 N and direction classification accuracy over 95 %. In practical tests, the sensor accurately measures the magnitude and direction of contact forces in nine typical directions, with a range of 0.45–10.5 N, demonstrating good performance. Compared with conventional vision-based tactile methods, the proposed approach reduces the complexity of the light source design and simplifies image processing while maintaining high accuracy. The method provides accurate perception and demonstrates great potential for applications in tactile sensing.</p>