This study introduces a deep learning model designed to efficiently detect and classify intracranial hemorrhage (ICH) subtypes using non-contrast head CT images. ICH, which involves bleeding within the skull, is a critical condition that necessitates quick and precise diagnosis. The hemorrhages are categorized into intra-axial (intraventricular and intraparenchymal) and extra-axial (subdural, epidural, and subarachnoid) based on their location. Previous computer-aided diagnosis (CAD) systems for ICH detection and classification typically focus on binary classification and have a high number of parameters, leading to increased storage requirements. Moreover, these models often lack the accuracy required for critical medical applications. Therefore, there is a need for a more efficient and accurate automated ICH detection system. To address these limitations, we developed a double- branch model based on the Xception architecture. This model extracts both spatial and temporal features, combines them, and generates a 3D spatial context. These combined features are then fed into a decision tree classifier for final predictions. The dataset for this study was obtained from the 2019 Radiologist Society of North America (RSNA) brain hemorrhage detection challenge. Our model surpassed existing benchmark models, demonstrating higher accuracy rates in detecting various hemorrhage types: intraventricular (96.59%), subarachnoid (96.59%), intraparenchymal (95.36%), and subdural (94.05%), Epidural (99.56%). These results confirm the effectiveness of the proposed double-branch Xception architecture in the detection and classification of ICH.