To address the challenges of limited generalization in 2D feature extraction and size measurement distortion caused by the loss of depth information in monocular vision during wide depth-of-field experimental observations of gas-liquid two-phase bubbly flows, this paper proposes a high-precision bubble size measurement method driven jointly by binocular vision and deep learning. For 2D bubble segmentation, the method integrates the VMamba backbone with a PAFPN neck and incorporates an occlusion-aware contour reconstruction module to construct an instance segmentation model named VM-BubNet, effectively resolving missed detections and incomplete masks caused by overlapping bubbles under real working conditions. For 3D spatial mapping, a pixel-to-physical space mapping model is established by combining the binocular vision system with epipolar geometric constraints, and a multi-dimensional feature-weighted matching algorithm is employed to achieve precise cross-view bubble pairing, enabling accurate computation of 3D spatial coordinates and subsequent size calibration. Experimental results demonstrate that VM-BubNet achieves a precision of 0.982 7 and a recall of 0.967 2 on complex experimental datasets, with the core segmentation metric Mask AP@0.5 reaching 0.973 6. The occlusion reconstruction mechanism effectively corrects systematic area errors induced by incomplete masks, and the proposed multi-dimensional feature matching algorithm consistently maintains an average matching rate exceeding 90% across various working conditions. Quantitative analysis based on calibrated bubble sizes further reveals the influence of hydrodynamic conditions on bubble evolution. Within the experimental range of gas flow rates from 0.06 to 0.24 m3/h and liquid flow rates from 9 to 12 m3/h, increases in both flow rates lead to a significant increase in the total number of bubbles in the flow field. Specifically, an increase in gas flow rate directly raises the initial bubble count, whereas an increase in liquid flow rate induces intense turbulent dissipation that continuously exacerbates secondary bubble breakage and causes a sharp surge in micro-bubbles. Compared with the gas flow rate, the liquid flow rate plays a dominant role in determining bubble size distribution characteristics. The resulting strong shear forces effectively suppress bubble coalescence and growth while significantly broadening the range of bubble size distribution. This method overcomes the limitations of traditional monocular detection schemes and general-purpose segmentation algorithms, providing a reliable technical solution for intelligent bubble recognition and quantitative 3D parameter analysis in industrial multiphase real-flow scenarios.