Abstract: Trichromatic mask particle image velocimetry employs optical elements to alter imaging paths, integrating color and perspective information for single-camera three-dimensional flow velocity measurements. Due to the small imaging viewing angle between three perspectives, conventional algebraic reconstruction techniques yield elongated particle distributions and low directional resolution along the imaging depth direction. To mitigate elongation effects and enhance three-dimensional particle reconstruction quality, convolutional neural networks (CNNs) are applied. Utilizing the AIPR (Artificial Intelligence Particle Reconstruction) architecture, an algorithm tailored for small viewing angle in trichromatic mask PIV is developed. Evaluation involves synthetic image data and artificial flow fields varying in different viewing angles and particle concentration. Results demonstrate that compared to traditional algebraic methods, the proposed method improves reconstruction quality by 30% at the particle concentration of 0.05 ppp. Additionally, it achieves a reconstruction speed of up to 2.8 times faster than traditional methods, enabling the rapid three-dimensional reconstruction of the particle field and enhancing axial reconstruction quality.