{"id":"adc862c6-89ed-43ec-b17a-311239347ecb","arxiv_id":"1908.09401","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A U-net reconstructs MNIST, EMNIST, and Kanji49 images from raw sensor data of a see-through lensless camera, but classification benefits are inconsistent and the manuscript is incomplete.","lead":"Researchers show that a convolutional neural network, specifically a U-net, can reconstruct images captured by a lensless see-through camera made of a transparent window and a bare CMOS sensor. The preprint is an unfinished draft, so the broader claim about comparable classification accuracy is not yet established.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported results do not establish a learned inverse of the see-through camera; they are consistent with memorizing a closed label set, so the central reconstruction claim is unsupported.","rationale":"The reader's weakest assumption correctly identifies the load-bearing issue: the closed-set, single-geometry, single-sensor design cannot separate true inverse learning from class memorization. The large training-to-test MAE gaps and the Kanji49 classification reversal are direct empirical symptoms of this failure. I agree with the REJECT verdict because the central reconstruction claim, as generalized in the abstract, is not supported by the data. The manuscript is also unfinished: the conclusion reads '...TO DO', and there are duplicated and garbled passages, including repeated dataset descriptions and duplicated tables/figures in the full text. These are not mere style issues; they make the experimental reporting unreliable even before considering the generalization concern. No code, data, or detail sufficient to reproduce the sensor-to-image pairs is provided, so the reported numbers cannot be independently checked. The proposed held-out-class experiment would settle whether the network has learned a general inverse or a label-set-specific lookup; until that test is reported, the central claim should be treated as unverified. This does not impugn the underlying idea, but the present draft does not meet the evidentiary standard for acceptance.","tokens_in":6906,"tokens_out":3860,"duration_ms":42517,"concrete_test":"Train the identical U-net on MNIST digits 0-5 captured with the same see-through camera geometry, then evaluate reconstruction on held-out MNIST digits 6-9 captured under identical conditions. If the network has learned the inverse of the space-variant PSF, test MAE on held-out classes should be comparable to the reported in-distribution test MAE (0.1014) and reconstructed digits should remain recognizable. If the network has learned class templates, held-out MAE will be substantially worse and reconstructions will tend to collapse to the trained classes. Report per-class MAE and a confusion matrix to confirm that held-out digits are not being mapped to training classes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that a U-net learns an inverse mapping from the see-through camera's space-variant sensor pattern to the image is underdetermined by the reported experiments. All training and test images come from the same closed label set (e.g., 6 MNIST classes) displayed on the same LCD at the same ~250 mm geometry. In this setting, a sufficiently expressive U-net can succeed by learning class-specific templates or a class-conditional lookup rather than a general inverse of the point-spread function. Table 1's MNIST training MAE of 0.0129 versus test MAE of 0.1014 (nearly 8x) and the reconstructed-image classifier's drop from 99.7% training accuracy to 80.9% testing accuracy indicate substantial overfitting. The Kanji49 results are even more direct evidence: raw-sensor test classification (43.3%) exceeds reconstruction-based test classification (30.8%), which contradicts the abstract's statement that similar classification accuracy is likely for raw versus reconstructed data. Therefore, the evidence does not distinguish genuine inverse learning from dataset-specific memorization, and the headline result may not transfer to unseen symbols, geometries, or sensors.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript describes a computational \"see-through\" camera, consisting of a transparent plexiglass window and a bare CMOS sensor, and proposes to reconstruct displayed images from the raw sensor measurements with a U-Net-style convolutional network. The authors report reconstruction MAE and classification accuracy for three datasets (MNIST restricted to 6 classes, EMNIST with 47 classes, and Kanji49 with 49 classes), comparing classification on original images, raw sensor images, and reconstructed images. The paper's central claim, stated in the conclusion, is that a U-net-based network can reconstruct MNIST images with good fidelity, and the abstract generalizes this to \"image reconstruction can be achieved via a convolutional neural network for a see-through computational camera.\" The manuscript is presented as a preliminary proof of concept rather than a full system demonstration.","tokens_in":7123,"tokens_out":1523,"duration_ms":15539,"significance":"If the central claim were established, the paper would add a data point to the growing body of work on learned inverses for lensless and computational cameras, and it would provide a useful comparison between classification directly on raw sensor data and classification after learned reconstruction. The authors are honest about the preliminary nature of the work and use publicly available datasets. However, the significance is substantially undercut by the incomplete state of the manuscript and by the lack of evidence that the learned reconstruction generalizes beyond the closed label sets used in training. The paper explicitly notes the absence of a completed conclusions section and contains duplicated and garbled passages, which makes it unsuitable for publication in its present form.","major_comments":[{"comment":"The manuscript is an incomplete draft: Section 5 is literally the placeholder text \"...TO D O\" and the reference list is cut off mid-list. This is not a presentation issue that can be fixed with light editing; the paper lacks a proper conclusion and a complete reference list, and several paragraphs (e.g., the duplicated dataset description in Section 3 and the duplicated figure captions) are garbled. A journal submission must be a complete document.","section":"Section 5 (Conclusions)"},{"comment":"The central generalization claim is underdetermined by the reported experiments. All training and test images come from closed label sets (e.g., 6 MNIST classes) displayed on the same LCD at the same ~250 mm geometry. The observed training-to-test MAE gap is large (MNIST: 0.0129 training vs. 0.1014 testing, nearly 8x), and the reconstructed-image classifier drops from 99.7% training accuracy to 80.9% testing accuracy. These gaps are consistent with the network learning label-set-specific templates rather than a general inverse of the camera's space-variant PSF. The paper provides no out-of-distribution test (e.g., unseen classes, different geometry, or different sensor) that would distinguish these hypotheses, so the central claim that a U-net learns a general inverse mapping is not supported.","section":"Section 3, Table 1 and Figure 6"},{"comment":"The Kanji49 results directly contradict the abstract's claim that \"similar classification accuracy is likely possible in both cases.\" For Kanji49, raw-sensor testing accuracy (43.33%) is higher than reconstructed-image testing accuracy (30.77%), a substantial and unexplained reversal. The paper's explanation (\"further parameter tuning\") is unfalsifiable as stated and does not address the fact that reconstruction is demonstrably hurting classification on this dataset. This finding weakens the paper's comparative claim about raw versus reconstructed classification.","section":"Section 4.2 / Figure 6 and Table 2"},{"comment":"The empirical results are reported without error bars, repeated trials, or any measure of variability. Table 1 and Figure 6 report single runs with no indication of the number of seeds, no standard deviation, and no statistical comparison between methods. Given the small generalization gap for some datasets and the large gap for others, the authors should report variance across independent training runs or justify why a single run is sufficient.","section":"Section 2 (Network and Training) and Section 3"},{"comment":"The comparison between raw-sensor classification and reconstructed-image classification is confounded by different input preprocessing: raw sensor images are resized to 125x170 pixels, while reconstructed images are resized to 32x32 pixels. This is a 22x difference in pixel count and may independently affect classifier performance, making the raw-versus-reconstructed comparison difficult to interpret. The paper does not discuss this confound.","section":"Section 4.2 / raw sensor resolution"}],"minor_comments":[{"comment":"The caption says \"The letter 'o' from the MNIST dataset\" but MNIST contains digits, not letters; this is likely a typo for a handwritten digit.","section":"Section 1, Fig. 1 caption"},{"comment":"The pixel-wise cross-entropy loss is introduced as a sparsity-inducing loss for binary pixels, but the ground-truth images are grayscale, not binary. The manuscript should clarify whether the ground truth is binarized before training or how the cross-entropy is applied to continuous intensities.","section":"Section 2, loss function"},{"comment":"The paragraph beginning \"For each data set, we trained and tested 3 classification networks...\" appears twice with slightly different wording, and the garbled section \"EM NI ST is the an augm entation of M NI ST...\" is a corrupted duplicate of the dataset description. These should be cleaned up.","section":"Section 3, duplicated text"},{"comment":"The reference list is incomplete; it ends at reference [9] despite references numbered up to [20] in the text, and some citations (e.g., ref. [15]) appear to be incorrectly numbered relative to the in-text callouts.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is not publishable in its current state: it is an incomplete draft with a missing conclusions section, garbled duplicated text, and an incomplete reference list. Beyond the mechanical issues, the scientific claim is not yet supported because the experiments do not rule out label-set-specific memorization as the mechanism behind the reported reconstructions. The authors should first complete the manuscript and then design an out-of-distribution test (e.g., reconstructing classes excluded from training, or varying the object-to-sensor distance) to demonstrate genuine inverse learning. I would not recommend rejection on grounds of novelty or scope, as the problem is relevant to the journal, but the current draft is far from a complete submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. First, this is an unfinished draft: the conclusion section literally reads '...TO D O', there are duplicated and garbled passages, and figures appear twice in different formatting. Sending that to reviewers would waste their time. Second, the evidence in the draft does not support the abstract's broad claim that a CNN learns to reconstruct images from the see-through camera. The numbers point toward the network memorizing the closed label set rather than learning a general inverse.\n\nThe new bit is real: previous work from this group showed classification from a lensless camera and introduced the see-through geometry; here they add a U-net-based reconstruction and directly compare classification from raw sensor data versus reconstructed images. That comparison is the most interesting part of the paper. The EMNIST result, where reconstruction lifts testing accuracy from about 2% to about 40%, is a concrete and non-obvious finding, even if still low.\n\nBut the manuscript is not close to publishable. No error bars, no repeated trials, no code or data. More substantively, the reconstruction results are consistent with class-conditional lookup. MNIST training MAE is 0.0129 versus 0.1014 on test, and the reconstructed-image classifier drops from 99.7% to 80.9%. For Kanji49, classifying raw sensor data beats classifying reconstructed images, 43.3% versus 30.8%, directly contradicting the abstract's suggestion that similar accuracy is likely. No out-of-distribution test is reported; every image comes from the same LCD, same geometry, same closed symbol set. So the paper does not establish that the U-net learned a true inverse of a space-variant PSF. The citation pattern is fine; self-citations to the group's prior work are appropriate.\n\nWho this is for: readers interested in lensless imaging and learned reconstruction might find the raw-vs-reconstructed comparison a useful preliminary data point, but only after the paper is completed and the overfitting question addressed. As a preprint, it is a draft.\n\nRecommendation: desk reject. The idea could merit a serious referee after a full rewrite, proper experimental reporting, and tests that separate generalization from memorization. Right now the paper is not ready.","headline":"Unfinished draft whose central reconstruction claim is undermined by the reported generalization gap; the raw-vs-reconstructed comparison is worth a look only after major completion.","tokens_in":7639,"tokens_out":2505,"would_cite":false,"duration_ms":22739,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A convolutional network trained on sensor patterns from a see-through, lensless camera can reconstruct the original displayed characters with good fidelity for simple datasets, and classification from those reconstructed images can rival…","keywords":["see-through camera","lensless imaging","computational imaging","image reconstruction","U-net","convolutional neural network","space-variant point-spread function","MNIST"],"falsifier":"Present the trained reconstructor with a displayed image it never trained on—for example, MNIST digits 6 through 9 if training used only digits 0 through 5—at the same geometry and sensor. If those unseen digits come out unrecognizable and the mean absolute error jumps far above the reported MNIST test value of 0.1014, the network has memorized the training classes rather than learned the camera's inverse; if unseen digits still reconstruct, the general-inverse claim survives.","tokens_in":6697,"feed_emoji":"📷","tokens_out":9902,"duration_ms":90229,"temperature":0.7,"pith_summary":"This paper asks whether a lensless 'see-through' camera—an image sensor mounted at the edge of a transparent window with no lens—can have its images recovered by a trained neural network rather than by a calibrated matrix inversion. The authors train a U-net-style convolutional network on sensor patterns produced by an LCD placed about 250 mm away, for three handwritten-character datasets, and measure both reconstruction error and downstream classification accuracy. The main claim is that for a restricted label set such as MNIST the network reconstructs the original images with good fidelity, and that classification from reconstructed images can match or beat classification from raw sensor data on simple sets. They also report that this advantage does not extend automatically to larger, more varied datasets such as Kanji49, where reconstructed-image classification is worse than raw-sensor classification.","feed_headline":"Neural network reconstructs images from see-through camera","feed_subtitle":"A U-net learns the inverse blur of a transparent-window lensless camera on MNIST, EMNIST, and Kanji49.","key_machinery":"The load-bearing mechanism is the U-net encoder-decoder: convolutional stages halve the spatial dimensions while doubling the channel count, decoder stages reverse that process, and skip connections pass encoder features to the decoder so fine detail is not lost. The network is augmented with dense blocks and trained with pixel-wise cross-entropy loss on sigmoid outputs, a choice that encourages sparse reconstructions. Because the camera's point-spread function is space variant—the blur at the sensor depends on where the object pixel came from—no single convolution kernel can undo the blur, so the network must learn a many-to-one mapping from sensor patterns back to displayed images. The classification comparison uses the same off-the-shelf classifier on raw sensor data and on reconstructed images, testing whether the reconstructed image is a useful intermediate representation rather than just a visually cleaner picture.","core_discovery":"On the paper's own terms, the finding is that a U-net-based convolutional network can learn an inverse mapping for a lensless camera whose blur changes across the field of view: for the six MNIST classes the training mean absolute error is 0.0129 and the testing error 0.1014; for the 47-class EMNIST set the values are 0.0730 and 0.1213; for the 49-class Kanji49 set they are 0.0994 and 0.1786. The classification comparison shows that for MNIST, classifying reconstructed images gives a testing accuracy of 0.809, above the 0.612 obtained from raw sensor data; for EMNIST, reconstruction lifts testing accuracy from 0.019 to 0.397; and for Kanji49, it drops from 0.433 to 0.308. The authors conclude that similar classification accuracy is likely achievable with either route given further network optimization, and that reconstruction is demonstrated with good fidelity for MNIST in particular.","pith_inferences":["A direct test of whether the network learned a true inverse rather than a label-set lookup would be to keep the camera fixed and reconstruct an image unlike anything in training, such as a handwritten digit from a class excluded from training or a photograph of a non-character.","The pattern across the three datasets suggests that as the label set grows, end-to-end classification on raw sensor data may be the more scalable route, with reconstruction useful mainly when humans need to see the image.","Because the camera blur is space variant, the trained network is likely tied to the exact window-sensor geometry; a testable extension is to train at one object distance and evaluate at another to see how far the learned inverse generalizes."],"forward_implications":["On a fixed display-sensor geometry, a see-through camera can act as an image-forming device for a restricted set of characters, turning a transparent window into a lensless display reader.","Classifying reconstructed images instead of raw sensor patterns can improve accuracy on simple label sets—MNIST testing accuracy rises from 0.612 to 0.809—so reconstruction can be a useful preprocessing step when the label space is small.","The benefit of reconstruction disappears on the largest, most variable dataset tested (Kanji49: reconstructed 0.308 vs raw 0.433), so reconstruct-then-classify is not generally superior to direct classification from sensor data.","Any practical deployment inherits the paper's fixed conditions: same LCD, same 250-mm geometry, same sensor, and same alphabets; changing any of these would require retraining or further experiments."],"supporting_citations":[{"why":"Previous demonstration that machine learning can classify images from lensless cameras, the direct predecessor this paper extends to reconstruction.","marker":"[9]"},{"why":"Introduces the see-through lensless camera setup (transparent window plus edge-mounted sensor) that all experiments use.","marker":"[10]"},{"why":"Provides the U-net encoder-decoder architecture with skip connections that the reconstruction network is built on.","marker":"[12]"},{"why":"Provides the residual-learning idea used to add dense blocks to the U-net for deeper training.","marker":"[13]"},{"why":"Supports the choice of pixel-wise cross-entropy loss for sparse image reconstruction.","marker":"[15]"},{"why":"Supplies the Adam optimizer used to train all reconstruction and classification networks.","marker":"[16]"},{"why":"Supplies the off-the-shelf classifier used to compare raw-sensor and reconstructed-image classification.","marker":"[17]"},{"why":"Supplies the MNIST dataset, the primary proof-of-concept with six classes.","marker":"[18]"},{"why":"Supplies the EMNIST dataset with 47 classes used to test a larger label set.","marker":"[19]"},{"why":"Supplies the Kanji49 Hiragana dataset used to test the most complex and variable label set.","marker":"[20]"}],"fun_headline_variants":["U-net learns inverse blur for see-through camera","See-through camera: deep learning turns blur into images","CNN reconstructs see-through camera images, boosts some classifications","Transparent-window camera: ML image reconstruction and classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central assumption is that a network trained and tested only on the three label sets, all displayed on the same LCD at the same 250 mm distance and captured by the same sensor, has learned a general inverse of the camera's blur rather than a per-class lookup.","fun_headline_variants_meta":{"raw":{"variants":["U-net learns inverse blur for see-through camera","See-through camera: deep learning turns blur into images","CNN reconstructs see-through camera images, boosts some classifications","Transparent-window camera: ML image reconstruction and classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000856,"raw_usage":{"total_tokens":3660,"prompt_tokens":826,"completion_tokens":2834,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":2770}},"tokens_in":442,"tokens_out":2834,"duration_ms":27164,"temperature":1.0,"reasoning_tokens":2770,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:12:02.112974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Present the trained reconstructor with a displayed image it never trained on—for example, MNIST digits 6 through 9 if training used only digits 0 through 5—at the same geometry and sensor. If those unseen digits come out unrecognizable and the mean absolute error jumps far above the reported MNIST test value of 0.1014, the network has memorized the training classes rather than learned the camera's inverse; if unseen digits still reconstruct, the general-inverse claim survives.","supporting_citations":[{"cited_title":"Lensless-camera based machine learning for image classification","cited_arxiv_id":"1709.00408","evidence_quote":"Previous demonstration that machine learning can classify images from lensless cameras, the direct predecessor this paper extends to reconstruction."},{"cited_title":"Computational imaging enables a “see-through","cited_arxiv_id":null,"evidence_quote":"Introduces the see-through lensless camera setup (transparent window plus edge-mounted sensor) that all experiments use."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the U-net encoder-decoder architecture with skip connections that the reconstruction network is built on."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the residual-learning idea used to add dense blocks to the U-net for deeper training."},{"cited_title":"Risk-sensitive loss functions for sparse multi-category classification problems,","cited_arxiv_id":null,"evidence_quote":"Supports the choice of pixel-wise cross-entropy loss for sparse image reconstruction."},{"cited_title":"LeCun and C","cited_arxiv_id":null,"evidence_quote":"Supplies the MNIST dataset, the primary proof-of-concept with six classes."}],"review_version":1}