<p>Limited-view optoacoustic tomography presents significant challenges in image reconstruction, especially in scenarios where high-quality full-view image is unavailable. To address this, we propose a reference-free sparse cooperative network that leverages twin implicit neural sub-networks to reconstruct high-quality images from complementary and sparsely noisy data, defined in an orthogonal decomposition of the image space. In the sparse measurement space and the its pseudo-inverse space, the input is encoded as spatially distributed latent representation by utilizing recurrent structures to enhance global feature modeling and capture long-range dependencies. The desired reconstructed photoacoustic image is regarded as an implicit continuous function over the 2D spatial domain, where pixel values are modeled as the output of a continuous mapping from spatial coordinates and prior image features. We design a self-supervised loss function for adaptively learning continuous and high-quality representation with three terms: the estimate of the mean squared error without the high-quality ground-truth data; total variation loss to enhance image continuity and promote smoothness in the generated images; the physics-consistency loss to restore full-angle coverage in the frequency domain. Our method consistently outperforms 8 state-of-the-art unsupervised approaches across 12 datasets, yielding up to 12.9% higher PSNR, 9.6% relative improvement in SSIM. Compared to supervised approach, our unsupervised paradigm exhibits more stable and efficiency convergence for optoacoustic imaging. Additionally, through quantitative evaluation metrics, spectral deviation analysis, visualization of neural tangent kernels, and loss function analysis, the results confirm that the proposed method allows for effective image reconstruction without the availability of large-scale high-quality ground-truth supervision.</p>