{"id":"a14fbfaa-75f0-42ec-88f2-5d72af4739c8","arxiv_id":"2505.01755","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An end-to-end network with a learnable Coded Mask Simulator and Wiener fusion improves lensless image reconstruction on two benchmark datasets.","lead":"LensNet trains a neural network to reconstruct images from lensless cameras by learning an approximate point-spread function and applying Wiener filtering in the frequency domain. It reports higher PSNR, SSIM, and lower LPIPS than several prior methods on the DiffuserCam and MWDNs benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 lacks any baseline-retraining protocol or error bars; without confirmation that all methods share the same split, resolution, and training loss, the central 'consistently outperforms' claim is not established.","rationale":"Reading the paper in good faith, LensNet appears to be a plausible end-to-end reconstruction network, and the ablation supports the contribution of its components. The reader's verdict of CONDITIONAL is appropriate. However, the most load-bearing concern about the central empirical claim is not the CMS-to-PSF ambiguity, which affects interpretation but not necessarily the reported numbers; it is the absence of any stated baseline-retraining protocol or error bars. Without knowing whether the competing methods were trained and evaluated under identical conditions, the 1.7–1.5 dB PSNR gaps could be artifacts of split, preprocessing, or loss choices. This is a standard correctness risk for an empirical SOTA claim. The concrete test of retraining all baselines under the same protocol would settle this concern. If the margins persist, the paper should be accepted; if they shrink, the claim needs to be softened. The CMS issue should also be addressed by specifying how the channel-attention output becomes the PSF in Eq. (4), but it is a secondary, interpretability-level concern. No formal verification or reproducible code was available to independently check the results, which further supports keeping the verdict conditional rather than full acceptance.","tokens_in":15072,"tokens_out":9554,"duration_ms":90787,"concrete_test":"Retrain every learning-based baseline in Table 1 on the exact DiffuserCam and MWDNs splits used for LensNet, using the same augmentation, optimizer, resolution, and (where applicable) the same loss as LensNet; report PSNR, SSIM, and LPIPS as mean±std over at least 3 random seeds. If LensNet's margin over MWDN on either dataset falls below 1 dB or is within one standard deviation of the baselines, the state-of-the-art claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LensNet establishes new state-of-the-art results on DiffuserCam and MWDNs, supported solely by Table 1. The paper never states whether the listed learning-based baselines were retrained under the same protocol or whether their numbers were copied from prior papers. DiffuserCam alone has been used with different train/test splits in the literature (e.g., 19k/1k vs. the 24k/1k used here), and the reported preprocessing and loss functions differ across methods. LensNet's loss in Eq. (8) explicitly includes an LPIPS term, which can directly improve the LPIPS metric and alter training dynamics; if baselines were trained with MSE only, the comparison is not apples-to-apples. The SSIM and LPIPS gains are therefore partly a function of the objective, not necessarily the architecture. The ablation in Table 2 shows internal design choices matter, but it does not validate the external comparison. The reader's CMS-to-PSF concern is real and important for the physical-interpretation claim, but it is secondary: even if the CMS output is not a true PSF, LensNet could still achieve the reported PSNR. The load-bearing gap for the SOTA claim is the missing experimental protocol for baselines, which is the least secure link between the experiments and the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"LensNet is an end-to-end deep learning framework for lensless image reconstruction. It combines an encoder-decoder architecture with a learnable Coded Mask Simulator (CMS) intended to estimate the point spread function from data, and Wiener Fusion Blocks that perform frequency-domain deconvolution. The model is trained with a weighted sum of MSE, SSIM, and LPIPS losses. Experiments on the DiffuserCam and MWDNs datasets compare LensNet with classical and learning-based baselines and report PSNR, SSIM, and LPIPS, together with an ablation study, a user study, and a model-complexity analysis.","tokens_in":15328,"tokens_out":3586,"duration_ms":36847,"significance":"If the reported results are reproducible under the claimed protocol, LensNet would offer a practically useful combination of learned PSF modeling and Wiener filtering for lensless imaging. The paper's strengths include a released code link, a broad set of comparisons, an explicit multi-term loss, and an internal ablation study. However, the central state-of-the-art claim rests on a single table whose baseline protocol is unspecified and whose numerical results are reported without measures of variance. In addition, the manuscript does not specify how the CMS output becomes the PSF used in the Wiener filter, which prevents verification of the physics-based modeling claim. These issues are addressable but currently block full confidence in the headline result.","major_comments":[{"comment":"The paper does not state whether the learning-based baselines (TikNet, FlatNet, LenslessGAN, UDN, MWDN) were retrained under the same protocol or whether their numbers were copied from prior publications. Because Eq. (8) trains LensNet with SSIM and LPIPS terms in addition to MSE, and Table 1 reports exactly those three metrics, the comparison is only meaningful if all baselines were trained with the same loss function, data split, image resolution, and preprocessing. Please specify the training protocol for every baseline, including the loss used, the train/test split, and the number of training epochs.","section":"§4.4, Table 1"},{"comment":"The CMS is described as outputting channel attention weights s_c ∈ R^{N×C′×1×1} after global average pooling and a 1×1 convolution, but the paper never explains how this vector-valued attention output is converted into the two-dimensional PSF(x,y) that enters the Wiener transfer function H(u,v) in Eq. (4). Without an explicit formula or an architectural description connecting the CMS output to PSF(x,y) and to the feature map B(u,v) being filtered, the claim that the CMS performs learnable PSF estimation cannot be verified or reproduced.","section":"§3.3, Eq. (4)"},{"comment":"No error bars, confidence intervals, or statistical significance tests are reported for any quantitative result. Table 1's improvements over the strongest baseline (e.g., PSNR 27.46 vs. 25.74 on DiffuserCam) may be real, but single-run numbers do not establish that LensNet 'consistently outperforms' all competing methods. The same issue applies to the ablation results in Table 2 and the user-study ratings in Table 3, where no measure of inter-participant variability or significance is provided.","section":"§4.3, Table 1; §4.5, Table 2"}],"minor_comments":[{"comment":"The ablation study does not state which dataset it was performed on, and the variants 'ThreeDown' and 'w original PSF' are not defined precisely. Please specify the dataset, the training protocol, and exactly what 'original PSF' means in the 'w original PSF' row.","section":"§4.5, Table 2"},{"comment":"The text says σ is a '(linear) activation function'; if the activation is linear, then the channel attention operation is affine and the description should say so explicitly and name the activation actually used in the experiments.","section":"§3.3, Eq. (7)"},{"comment":"Figure 3 contains labels such as '???' and 'Random Matrix' that are not explained in the caption or in Section 3.2. Please clarify the notation and the role of the random matrix and learnable parameters in the diagram.","section":"Figure 3"},{"comment":"The DiffuserCam dataset split of 24,000 training and 1,000 test images should be justified with respect to the splits used by the cited baselines, since different papers in the literature use different train/test partitions for this dataset.","section":"§4.1"},{"comment":"The abstract and conclusion state that LensNet 'establishes a new standard' and 'outperforms state-of-the-art methods'; these claims are stronger than what a single dataset pair and an unspecified baseline protocol can support. Phrasing such as 'outperforms the compared methods under the reported protocol' would be more appropriate.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is plausible and the experimental section is extensive, but the missing baseline training protocol and the unspecified CMS-to-PSF mapping are load-bearing issues for the main claims. The authors should be asked to provide full protocol details and, ideally, variance estimates or at least multiple seeds. The user study also needs a statistical analysis before it can support the 'consistently outperforms' wording."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: LensNet is a competent empirical deep-learning paper for lensless reconstruction. The architecture—encoder-decoder with a learnable Coded Mask Simulator, a Wiener fusion block, and multi-scale spatial amplification—is a genuine combination I haven't seen in prior work, and the reported numbers on DiffuserCam and MWDN are impressive. The ablation in Table 2 shows the components matter, which is the right kind of evidence. If the numbers are real and reproducible, this is a useful contribution to the field.\n\nThat said, the paper as written does not establish the SOTA claim, and two gaps need attention. First, Table 1 gives no baseline protocol. The paper doesn't say whether the competing methods were retrained under the same split and resolution or whether the numbers come from previous publications. DiffuserCam has been used with different splits in the literature, and LensNet's loss includes an LPIPS term, while baselines may have been trained with MSE or something else. That makes the PSNR/LPIPS comparisons potentially biased in LensNet's favor. There are also no error bars or significance tests, so the small margins on MWDN (e.g., SSIM 0.960 vs. 0.957) could be noise. The stress-test note is right that this is the weakest link.\n\nSecond, the CMS-to-PSF step is under-specified. Section 3.3 describes a channel-attention mechanism that 'implicitly recovers' the PSF, but it never states how the attention output becomes the PSF(u,v) used in Eq. (4). If the CMS output is just a feature map with no explicit role in the Wiener filter, then the physics-based framing is misleading, and the method reduces to a regular image-to-image network. That doesn't kill the empirical contribution, but it does require a correction in claims.\n\nMinor stuff: the ablation is only on MWDN, the user study has no error bars, and training details (learning rate, epochs) are vague. The math itself is standard and clean; I don't see any errors in Eqs. (1)-(5).\n\nWho this is for: anyone working on lensless imaging or computational imaging who wants a strong new baseline. It deserves a serious referee—the architecture is plausible, the experiments are extensive, and the weaknesses are fixable. I'd send it to peer review with major revision, requiring the authors to specify the baseline protocol, run multiple seeds with error bars, and either formalize the CMS-to-PSF connection or soften the 'physics-based' claim.","headline":"Solid empirical lensless-reconstruction paper with a genuine architecture, but the SOTA claim is not yet established because the baseline protocol is missing and the PSF mechanism is under-specified.","tokens_in":15872,"tokens_out":2953,"would_cite":false,"duration_ms":29266,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LensNet claims that an end-to-end network with a learnable Coded Mask Simulator and Wiener fusion blocks can reconstruct lensless camera images without fixed PSF calibration, reporting PSNR of 27.46 dB on DiffuserCam and 33.22 dB on MWDNs.","keywords":["lensless imaging","point spread function","Coded Mask Simulator","Wiener filtering","deep learning reconstruction","DiffuserCam","MWDNs","frequency-domain deconvolution"],"falsifier":"Take the learned distribution map from a trained LensNet, treat it as the PSF, and use Eq. (1) to synthesize a measurement from a known ground-truth image; then compare that synthetic measurement with the actual sensor measurement. If the synthesized measurements clearly differ, or if replacing the learned PSF with the measured or calibrated PSF changes reconstruction quality little, then the CMS is not modeling the true PSF and the claimed mechanism is not what drives the reported gains.","tokens_in":14851,"feed_emoji":"📷","tokens_out":13685,"duration_ms":109403,"temperature":0.7,"pith_summary":"LensNet claims that lensless image reconstruction improves when the point spread function (PSF) is learned from data rather than fixed by calibration. The paper proposes an end-to-end network that combines spatial-domain compression with frequency-domain Wiener deconvolution, using a Coded Mask Simulator to generate a data-driven PSF estimate during training. On the DiffuserCam and MWDNs benchmarks, LensNet reports PSNR of 27.46 dB and 33.22 dB respectively, beating the previous best learned method on both datasets, along with lower LPIPS values. If true, this means compact lens-free cameras could adapt to noise, misalignment, and changing scenes without explicit PSF calibration.","feed_headline":"LensNet learns the camera's blur and undoes it without PSF calibration","feed_subtitle":"It adapts to noise and misalignment without per-camera calibration, beating earlier lensless cameras on two benchmarks.","key_machinery":"The load-bearing component is the Coded Mask Simulator (CMS), a channel-attention module that pools the measurement feature map globally, passes it through a $1\\times1$ convolution with a linear activation, and outputs per-channel weights that are treated as the learned intensity distribution of the coded mask and hence of the PSF. That output feeds the Wiener Fusion Block (WNFB), which takes the fast Fourier transform of both the spatial features and the CMS output, applies the Wiener transfer function $H(u,v) = \\overline{\\mathrm{PSF}}(u,v)/(|\\mathrm{PSF}(u,v)|^2 + \\delta)$, and returns the restored signal to the spatial domain via the inverse FFT. The surrounding encoder-decoder structure, built from Reconstruction Blocks and Spatial Amplification Modules, fuses these spatial and frequency representations at multiple scales. The paper's argument is that this design lets the network adapt the PSF to the actual imaging conditions during training and deconvolve scattering and diffraction artifacts without hand-calibrated kernels.","core_discovery":"The central claim, stated on the paper's own terms, is that jointly estimating the optical encoding and inverting it in one network yields reconstructions that outperform both classical solvers and prior learning-based pipelines. LensNet treats the measurement as the convolution $I_{\\mathrm{measurement}} = I_{\\mathrm{object}} \\ast \\mathrm{PSF} + \\mathrm{noise}$, encodes the measurement with a Spatial Compression Module, and uses a Coded Mask Simulator to produce a learned distribution map intended to represent the coded mask's PSF. A Wiener Fusion Block then applies the transfer function $H(u,v) = \\overline{\\mathrm{PSF}}(u,v)/(|\\mathrm{PSF}(u,v)|^2 + \\delta)$ in the Fourier domain to restore the signal at multiple scales. With a combined MSE, SSIM, and LPIPS loss, the trained model reaches PSNR/SSIM/LPIPS of 27.46/0.863/0.099 on DiffuserCam and 33.22/0.960/0.024 on MWDNs, which the paper reports as consistently better than all compared methods. The authors interpret this as evidence that a learnable PSF, rather than a fixed or sparsely calibrated kernel, is the key to high-quality lensless reconstruction.","pith_inferences":["Because Section 3.3 does not define how attention weights become $\\mathrm{PSF}(x,y)$, the CMS output may functionally be an adaptive frequency filter rather than a faithful optical model; comparing the learned map with a measured PSF would settle which interpretation is correct.","If the physical PSF interpretation fails, LensNet can still be understood as an unrolled Wiener-deconvolution network with a learned regularizer, a weaker but still useful contribution for lensless imaging.","The same joint PSF-estimation-and-inversion idea could be tested on phase-mask cameras, where the PSF depends on wavelength; a per-channel CMS would be a natural extension.","The largest reported gains are on DiffuserCam, so a stress test under simulated PSF mismatch (for example, a slight rotation or shift of the mask) would show whether the learned PSF genuinely adapts or merely overfits the training distribution."],"forward_implications":["On DiffuserCam, LensNet reports PSNR 27.46 dB and SSIM 0.863, beating the prior best learned method MWDN (25.74 dB, 0.816) and lowering LPIPS from 0.132 to 0.099.","On MWDNs, LensNet reports PSNR 33.22 dB and LPIPS 0.024, again ahead of MWDN (31.74 dB, 0.030) and of all classical and learning-based baselines in the comparison.","A camera built on this approach would not need per-device PSF calibration or handcrafted preprocessing, because the network learns the PSF from data during training and can adapt to noise and system imperfections.","Reconstructions preserve high-frequency structure better than prior methods, as shown by qualitative comparisons that keep fur, hair, and fine textures sharper than FISTA, ADMM, FlatNet, UDN, and MWDN outputs.","The user study reported in the appendix gives LensNet the highest average ratings across reconstruction accuracy, detail restoration, artifact presence, noise handling, and visual realism among the methods tested."],"supporting_citations":[{"why":"Supplies the Wiener filter transfer function that LensNet embeds in its frequency-domain fusion blocks.","marker":"[Wiener, 1949]"},{"why":"Provides the linear convolution forward model that defines the lensless imaging problem.","marker":"[Goodman, 2005]"},{"why":"Introduces the DiffuserCam lensless camera and supplies the DiffuserCam dataset used for training and testing.","marker":"[Antipa et al., 2017]"},{"why":"Establishes the learned-reconstruction approach and the dataset preprocessing pipeline that LensNet builds on and compares against.","marker":"[Monakhova et al., 2019]"},{"why":"Provides the FlatNet and TikNet baselines, paired-training reconstruction methods whose high-frequency limitations LensNet aims to overcome.","marker":"[Khan et al., 2020]"},{"why":"Provides the LenslessGAN baseline and the competing idea of joint PSF estimation during reconstruction.","marker":"[Rego et al., 2021]"},{"why":"Provides the UDN unrolled primal-dual network baseline used in the quantitative comparisons.","marker":"[Kingshott et al., 2022]"},{"why":"Provides the MWDN baseline and the MWDNs dataset, the strongest prior method that LensNet claims to surpass.","marker":"[Li et al., 2023a]"},{"why":"Provides the FISTA baseline representing classical iterative optimization in the comparison table.","marker":"[Beck and Teboulle, 2009b]"}],"fun_headline_variants":["No more blur guessing: LensNet models and corrects lensless images in one pass","LensNet: end-to-end lensless imaging that learns its own PSF to beat calibration","Learn the blur, fix the image: LensNet outdoes fixed PSF methods","Lensless cameras get a brain: LensNet adapts PSF on the fly","From raw sensor to sharp image: LensNet's learned PSF wins on benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire gain rests on the assumption that the Coded Mask Simulator's channel-attention output is the physical point spread function used in the Wiener deconvolution step; Section 3.3 gives no explicit formula converting those attention weights or the distribution map into $\\mathrm{PSF}(x,y)$, so the improvement could come from the learned image-to-image mapping rather than from accurate PSF estimation.","fun_headline_variants_meta":{"raw":{"variants":["No more blur guessing: LensNet models and corrects lensless images in one pass","LensNet: end-to-end lensless imaging that learns its own PSF to beat calibration","Learn the blur, fix the image: LensNet outdoes fixed PSF methods","Lensless cameras get a brain: LensNet adapts PSF on the fly","From raw sensor to sharp image: LensNet's learned PSF wins on benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":4095,"prompt_tokens":1103,"completion_tokens":2992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":719,"completion_tokens_details":{"reasoning_tokens":2881}},"tokens_in":719,"tokens_out":2992,"duration_ms":20390,"temperature":1.0,"reasoning_tokens":2881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:10:47.365943+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the learned distribution map from a trained LensNet, treat it as the PSF, and use Eq. (1) to synthesize a measurement from a known ground-truth image; then compare that synthetic measurement with the actual sensor measurement. If the synthesized measurements clearly differ, or if replacing the learned PSF with the measured or calibrated PSF changes reconstruction quality little, then the CMS is not modeling the true PSF and the claimed mechanism is not what drives the reported gains.","supporting_citations":[{"cited_title":"Extrapolation, interpola- tion, and smoothing of stationary time series: with engi- neering applications","cited_arxiv_id":null,"evidence_quote":"Supplies the Wiener filter transfer function that LensNet embeds in its frequency-domain fusion blocks."},{"cited_title":"Introduction to Fourier optics","cited_arxiv_id":null,"evidence_quote":"Provides the linear convolution forward model that defines the lensless imaging problem."},{"cited_title":"Diffusercam: lensless single-exposure 3d imaging","cited_arxiv_id":null,"evidence_quote":"Introduces the DiffuserCam lensless camera and supplies the DiffuserCam dataset used for training and testing."},{"cited_title":"Learned reconstructions for practical mask-based lensless imaging","cited_arxiv_id":null,"evidence_quote":"Establishes the learned-reconstruction approach and the dataset preprocessing pipeline that LensNet builds on and compares against."},{"cited_title":"Flatnet: Towards photorealistic scene reconstruc- tion from lensless measurements","cited_arxiv_id":null,"evidence_quote":"Provides the FlatNet and TikNet baselines, paired-training reconstruction methods whose high-frequency limitations LensNet aims to overcome."},{"cited_title":"Robust lensless image reconstruction via psf estimation","cited_arxiv_id":null,"evidence_quote":"Provides the LenslessGAN baseline and the competing idea of joint PSF estimation during reconstruction."},{"cited_title":"Unrolled primal- dual networks for lensless cameras","cited_arxiv_id":null,"evidence_quote":"Provides the UDN unrolled primal-dual network baseline used in the quantitative comparisons."}],"review_version":1}