Amid the rapid development of the chemical industry, the large-scale production and use of hazardous chemicals have brought significant environmental risks and potential threats to human health. Consequently, there is an urgent need to develop efficient monitoring and identification methods to address these challenges. Mid-infrared (MIR) spectroscopy, characterized by its unique molecular “fingerprint” recognition capability and high sensitivity, has been widely applied in the identification and analysis of hazardous chemicals. Meanwhile, recent breakthroughs in deep learning, particularly in feature extraction and object detection, have provided novel solutions for addressing complex data analysis problems. To meet the need for precise identification of hazardous chemicals under complex background conditions, this study proposes a novel model that combines Depthwise Separable Convolution (DSC) and a High-Order Residual Network (HORN), referred to as DSC-HORN. This model is designed to enhance further the efficiency of feature extraction and the accuracy of recognition for complex MIR spectral data. By leveraging depthwise separable convolution, the DSC-HORN model decomposes standard convolutions into depthwise and pointwise convolutions, resulting in a reduction of approximately 65.89% in parameter count and computational complexity. This optimization increases the operational speed of the DSC-HORN model by approximately four times compared to traditional one-dimensional convolutional neural networks (1D-CNNs), while significantly reducing memory consumption and enhancing its adaptability in resource-constrained environments. For the experiments, this study collected MIR spectral data of a 10% DMMP solution mixed with 11 types of background materials. After data preprocessing and removal of outlier samples, a total of 528 valid spectral datasets were obtained for model construction. To ensure balanced distribution of different labeled samples across subsets and to accurately reflect the overall data characteristics, the dataset was divided into training, validation, and testing sets using a stratified random sampling method, with a 6∶1∶3 ratio for training, validation, and testing, respectively. The experimental results demonstrated that the DSC-HORN model exhibits significant advantages in recognition efficiency and accuracy compared to existing models. Compared to traditional models such as K-Nearest Neighbors(KNN), Random Forest(RF), Robust Mode Regression Sparse Non-negative Matrix Factorization(MR-NMF), and Backpropagation Neural Networks(BP), as well as deep learning models like1D-Convolutional Neural Network(1D-CNN), the DSC-HORN model achieved a recognition accuracy of 96.84%. This represents improvements of 5.7%, 13.93%, 5.07%, 10.76%, and 6.33%, respectively, over the models above. These performance enhancements can be attributed to the parameter optimization achieved through depthwise separable convolution and the efficient feature extraction enabled by high-order residual connections. The results further confirm that the DSC-HORN model is not only an efficient and accurate MIR hazardous chemical recognition model but also provides a novel technical pathway for real-time monitoring and precise identification of hazardous chemicals. Additionally, the MIR recognition technology, based on the improved high-order residual network, has great potential to promote further the widespread application of MIR spectroscopy in complex scenarios.