In the era of the Internet of Things, the exponential growth of interconnected devices has introduced new vulnerabilities and challenges for intrusion detection. Traditional packet-based detection methods often struggle to capture the subtle and sparse patterns inherent in modern attack traffic. This paper proposes TIGF-Net, a Traffic Image Gray Fusion Network that transforms network traffic sequences into gray-scale images and leverages multi-scale feature fusion to enhance detection robustness. The method employs an enhanced Multi-Scale Feature Fusion block specifically optimized for gray-scale representations, capturing both local and global structural patterns in attack and benign traffic. A Laplacian-based sharpening convolutional kernel is applied during preprocessing, which enhances high-frequency details. By incorporating edge-guided enhancement, key abnormal areas are highlighted. Extensive experiments on benchmark intrusion datasets demonstrate that the TIGF-Net approach exhibits superior performance in multi-level attack classification and adversarial robustness.