{"id":"5c58ccd4-67e8-4342-ba53-960139d23003","arxiv_id":"1909.05146","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Word and character segmentation is performed directly on run-length encoded handwritten document images, with 90.29% word and 86.23% character accuracy on ICDAR13 English handwritten data.","lead":"This paper presents methods for splitting handwritten document images into words and characters while the images remain in run-length compressed form, without decompressing them. A reader may care because processing compressed images can save time and memory for large document archives, though the reported accuracy is modest.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'min-cut' character segmentation is really a band-OR heuristic whose key assumption—that adjacent characters touch only in mid-bottom—is unmeasured; full-height or slanted joins defeat it.","rationale":"The reader's weakest assumption points at exactly the same vulnerability: character separation depends on the OR of top and bottom bands exposing gaps. I agree. The concern is load-bearing because the paper's own wording acknowledges it is a 'mostly' property and provides no data on its frequency. My stress-test adds that the paper's 'min-cut graph concept' is not a graph algorithm at all—no graph or cut objective is defined—so the only operative mechanism is gap detection in the OR band. Consequently, the reported 86.23% is conditional on an unmeasured property of the test data. This does not force a new verdict beyond the reader's CONDITIONAL: the method might work on the subset where the assumption holds, but the paper must quantify the assumption's coverage and compare against a real graph-cut baseline before the claim is accepted.","tokens_in":6428,"tokens_out":5553,"duration_ms":63749,"concrete_test":"Using the ICDAR13 ground truth, compute the top-bottom OR band at each ground-truth character boundary for every word in the test set (using the same ROI threshold as the paper). Measure the fraction of true boundaries at which the OR band has no gap; on those boundaries the algorithm cannot cut by construction. Then split boundaries into 'mid-bottom-only contact' vs 'full-height or slanted contact' by visual inspection and recompute character segmentation accuracy per stratum. If the full-height or slanted stratum recall is substantially below the overall 86.23%, the Section 3 assumption is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central character-segmentation claim rests on the assertion in Section 3 that 'the connection between the adjacent characters mostly occurs in the mid-bottom regions.' The algorithm then performs a logical OR of the top and bottom horizontal bands and treats gaps in that OR as minimum-cut edges (Figure 14, Algorithm Step 4). No graph or min-cut computation is actually specified: the cut point is simply the first x-interval in which the OR band has no foreground. If two characters touch in both the top and bottom bands—which happens in slanted or cursive English handwriting—the OR band is continuous across the boundary, no gap exists, and the method cannot place a character boundary. The paper gives no statistics on how often ICDAR13 word images satisfy the mid-bottom-only assumption, so the reported 86.23% cascaded accuracy is not anchored to the assumption actually holding. The word-gap threshold and the ROI threshold t are also set empirically with no formulas, so the result is a fitted value for these datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes methods for word and character segmentation performed directly on run-length encoded (RLE) handwritten document images. Word segmentation identifies connected components of foreground runs and thresholds inter-component gaps to group components into words. Character segmentation restricts analysis to a region of interest (ROI), divides it into top, middle, and bottom horizontal bands, performs a logical OR of the top and bottom bands to expose cut points between touching characters (described as a min-cut graph concept), and then applies insertion/deletion rules based on component lengths to address over- and under-segmentation. Experiments on ICDAR13 and several non-English datasets report 90.29% word segmentation accuracy and 86.23% cascaded character segmentation accuracy for English, with lower accuracies for Kannada, Oriya, Bangla, and Persian. The paper also claims lower computational complexity relative to pixel-domain processing and presents qualitative results on machine-printed images.","tokens_in":6604,"tokens_out":5215,"duration_ms":45620,"significance":"If the central claim holds, namely that word and character segmentation can be performed accurately and efficiently in the compressed domain, the work would be a useful contribution to document image analysis, especially for applications that need to avoid repeated decompression and recompression. The idea of exploiting run-length structure for segmentation is attractive, and the reported accuracies on ICDAR13 are non-trivial. The paper's strongest conceptual asset is the complexity argument for compressed-domain processing, which could motivate further work in this direction. However, the paper does not provide a direct comparison with existing segmentation methods, does not report sample sizes or error bars, does not make code or data available, and does not validate the key assumption about how adjacent characters touch. The thresholds are empirically selected from the test data, so the reported numbers are in-sample fits rather than predictions. Consequently, the significance of the claimed results cannot be fully assessed from the manuscript as it stands.","major_comments":[{"comment":"The algorithm described as a 'min-cut graph concept' is actually a band-OR heuristic: no graph, cut edges, or edge weights are defined, and the separation point is simply the first x-interval in which the logical OR of the top and bottom bands has no foreground. This mischaracterization affects the claimed novelty of using a graph-based min-cut. More importantly, the method's key assumption that adjacent characters connect only in the mid-bottom region is unvalidated: the paper does not report how often ICDAR13 character pairs satisfy this condition, so the 86.23% cascaded accuracy is not anchored to the assumption actually holding. Cursive or full-height/slanted joins would make the OR band continuous across the boundary and the method would fail. Provide statistics on touching-character geometry and compare against a pixel-domain segmentation baseline.","section":"Section 3, Algorithm Step 4 and Figure 14"},{"comment":"No sample sizes, number of test images, words, characters, error bars, or statistical tests are reported, and the matching criterion behind equations (7) and (8) is unspecified. The 90.29% word and 86.23% character accuracies are raw in-sample match rates on an unspecified subset of ICDAR13. Without a comparison to ICDAR13 contest participants or other published segmentation methods, the central claim that compressed-domain processing is competitive cannot be assessed. Additionally, the character accuracy is cascaded after word segmentation, so errors propagate; a standalone evaluation on ground-truth word segments is needed to isolate character-segmentation performance.","section":"Section 2.2, Tables 2 and 4; Section 3.2, Equation (8)"},{"comment":"Equations (2)-(6) are internally inconsistent and appear to treat the run-length representation as a binary image. Equation (2) defines CR(i) as a row sum over RLE(i,j) but does not define what RLE(i,j) represents for a run-length encoding, while equation (6) simultaneously uses RLE(i,j) as a binary pixel value and as a run-length term. As a result, the O(m x n) complexity claim is unsupported: if m and n are uncompressed image dimensions, the algorithm is not actually operating in the compressed domain; if they are compressed dimensions, the equations and complexity bound need a proper RLE formulation. This directly affects the paper's central premise that processing directly in compressed data saves time and memory.","section":"Section 2, Equations (2)-(6) and Section 2.1 complexity analysis"},{"comment":"Threshold selection is described inconsistently: Section 2 states a threshold is 'empirically chosen,' Algorithm Step 3 says 'Compute average gaps to automatically select a threshold value,' and Section 3.1 uses an empirically set ROI threshold t. If thresholds are computed from the test data themselves, the reported accuracies are in-sample fits rather than predictions. The paper must specify a training/test protocol or provide a sensitivity analysis over threshold values to show that the results are not due to dataset-specific tuning.","section":"Section 2.1 Algorithm Step 3 vs. Sections 2 and 3.1"}],"minor_comments":[{"comment":"Section 4 presents only two figures for machine-printed images with no quantitative evaluation; either provide numerical results or remove the claim that the model has been tested on machine-printed document images.","section":"Section 4"},{"comment":"The abstract and conclusion state that a 'min-cut graph concept' is used, but the body of the paper does not define a graph; please align the terminology with the actual OR-band procedure described in Section 3.","section":"Abstract and Conclusion"},{"comment":"The wording 'spread of every character' is misleading because the algorithm computes connected components of foreground runs, which may merge multiple characters; consider rewording to 'spread of connected components.'","section":"Section 2"},{"comment":"The 'one to one matching' used in equations (7) and (8) is not defined; specify the matching rule, including whether partial overlaps count and how ground-truth entities are matched to detected segments.","section":"Equations (7) and (8)"},{"comment":"Reference [Javed et al, 2016] and the later citation to 'Javed et al (Ph.D. Thesis), 2016' appear to refer to the same work; also, the datasets 'PBOK (Section 1.6.1)' and 'MARG dataset' are not described in the paper, so the reader cannot reproduce the experiments.","section":"References and Datasets"}],"recommendation":"major_revision","confidential_remarks":"The central idea is worth pursuing, but the manuscript substantially overclaims what is demonstrated. The 'min-cut graph concept' is not a graph-based method, the key assumption about character overlap is unmeasured, and the evaluation lacks the controlled comparisons and sample sizes expected for a claim of this scope. I would ask for a revised version with a proper evaluation protocol, a corrected description of the character-separation procedure, and a sensitivity analysis of the thresholds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Mohammed,\n\nThis is an incremental but genuinely new extension: the authors take their compressed-domain line segmentation and push it to word and character level for handwritten documents. Nobody else seems to have done that directly on run-length encoded data, so the niche is real. They also deserve credit for reporting performance on multiple non-English scripts and admitting the method degrades there.\n\nThe central approach is plausible: find connected components from foreground runs, use larger gaps to split words, and for character cuts, OR the top and bottom horizontal bands of the ROI and use the first x-gap in that OR. The complexity argument—O(mn) on the compressed representation versus O(hw) on the pixel grid—is fine and could matter for large archives.\n\nThe soft spots are substantial. The 'min-cut graph' is not a graph cut; it is a band-OR heuristic. It only works when adjacent characters touch in the mid-bottom region and leave a gap in the top and bottom bands. The paper asserts this is typical for English handwriting but supplies no measurement of how often ICDAR13 words satisfy it. In slanted or cursive hands, characters touch across the full height and the OR band is continuous, so the method silently fails. The reported 86.23% cascaded accuracy is therefore not anchored to a measured property; it is a fitted value on a specific dataset. Thresholds for the inter-word gap and ROI are hand-set, there are no error bars, no comparison against any decompress-then-segment baseline, and no code or data. I also could not follow equations (2)–(6): CR(i) is defined as a row sum, then used as if it were per-column coordinates, and the coordinate mapping is not reproducible.\n\nWho benefits: researchers working on compressed document image analysis, especially those building on Javed-style RLE processing. It is a useful existence proof, not a reliable method yet. I would send it to peer review if the venue tolerates heavy revision; the idea is worth one round of serious refereeing. A desk reject would also be defensible given the evaluation, but I think the uniqueness of the compressed-domain claim tips it toward review.","headline":"A genuinely new compressed-domain word/character segmentation idea, undermined by a mislabeled 'min-cut' heuristic and an evaluation too weak to support the accuracy claims.","tokens_in":7122,"tokens_out":2793,"would_cite":false,"duration_ms":28721,"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":"Handwritten word and character segmentation can be done directly on run-length compressed document images without decompressing to pixels.","keywords":["word segmentation","character segmentation","run-length encoding","compressed domain document image analysis","handwritten document images","connected components","min-cut graph","ICDAR13"],"falsifier":"Run the OR-of-bands character segmentation on a set of English handwritten words whose ground truth includes many full-height touching letters, such as tightly cursive 'th' or 'll' joins; if the OR pattern shows no gap at a large share of those true boundaries, the central assumption fails. A simpler check is to measure the cascaded character accuracy on a subset of ICDAR13 marked for heavy cursive versus a subset with separated letters; the method should perform systematically worse on the cursive subset.","tokens_in":6211,"feed_emoji":"✂️","tokens_out":6814,"duration_ms":66841,"temperature":0.7,"pith_summary":"This paper sets out to prove that word and character segmentation, a step normally run on decompressed pixel images, can be performed directly on the run-length encoded representation of a handwritten document image. The motivation is practical: if segmentation can run on the compressed form, document analysis pipelines avoid repeated decompression and recompression, saving both computation time and memory. The authors trace the horizontal spread of foreground runs to build connected components, split words by the large gaps between those components, and separate touching characters with a min-cut procedure on band-wise run patterns. They report 90.29% word segmentation accuracy and 86.23% cascaded character segmentation accuracy on the ICDAR13 English handwritten dataset, with lower accuracy when the same method is applied to non-English scripts.","feed_headline":"Compressed handwriting segmented at 90% word accuracy","feed_subtitle":"Run-length encoded scans are split into words and letters without decompression, saving time and memory for OCR pipelines.","key_machinery":"The machinery is the run-length representation of the image, together with a cumulative-run computation that turns each foreground run into a component spread [xmin, xmax]. Word segmentation uses an empirically chosen threshold on the gaps between these spreads to separate inter-word gaps from intra-word gaps. Character segmentation restricts attention to a region of interest, splits it into three horizontal bands, and performs a logical OR of the top and bottom bands; the gaps in the resulting pattern are treated as minimum cut edges where adjacent characters separate. Over- and under-segmentation are then repaired by inserting or deleting separator points according to component lengths relative to the average.","core_discovery":"The core claim is that the geometry of handwriting segmentation is visible directly in run-length encoded data: connected components can be established purely from the horizontal spreads of foreground runs, inter-word boundaries are the large gaps between those components, and within a word the touching of adjacent characters can be resolved by a logical OR of the top and bottom bands of the word's region of interest. The authors show that this is enough to reach 90.29% word accuracy and 86.23% cascaded character accuracy on ICDAR13, and they argue that working in the compressed domain costs O(m×n) time in the worst case, the same order as reading the compressed data itself, while pixel-domain processing would cost O(h×w). The character-segmentation step is explicitly built for English handwriting, where connections between neighboring letters occur mostly in the mid-bottom region.","pith_inferences":["A natural next step is to replace the empirically fixed gap threshold with a data-driven bimodality test on the component-gap histogram; because the paper already computes average gaps, a statistical split point would make the method parameter-free and easier to port across scripts.","The OR-of-bands procedure encodes a general principle: in any script where character joining is confined to a narrow vertical zone, logical combinations of band-wise run patterns will expose separators; this suggests a testable extension for Arabic, Devanagari, or Chinese by first locating the joining zone automatically.","Since the spread-based features are purely horizontal, the method's sensitivity to skew could be quantified directly: rotating a test line by small angles and measuring when word accuracy drops would give an explicit skew tolerance, which the paper leaves outside its scope."],"forward_implications":["OCR-style pipelines that already receive compressed scans can segment words and characters without decompressing to a pixel grid, removing the buffer and time cost of repeated decode-encode cycles.","Because the word boundary test is just a gap-length threshold, the method transfers to any script whose word spacing is consistently larger than intra-word spacing; the accuracy figures on Kannada, Bangla, Oriya, and Persian show the transfer is partial.","Character accuracy is cascaded: errors in word segmentation propagate directly into character segmentation, so a 90.29% word accuracy caps the ceiling for the 86.23% character accuracy.","The method requires no training and no pixel classification, only run-length data, which makes it a candidate for memory-constrained embedded document scanners."],"supporting_citations":[{"why":"Provides the compressed-domain text-line segmentation whose output feeds the word and character algorithms.","marker":"[Amarnath et al, 2018]"},{"why":"Establishes the prior compressed-domain word and character segmentation for machine-printed images that this work extends to handwriting.","marker":"[Javed et al, 2016]"},{"why":"Supplies the ICDAR13 handwriting segmentation contest data and ground-truth matching used to report accuracy.","marker":"[Nikolaos et al, 2013]"},{"why":"Makes the case that handwritten script segmentation is a critical challenge for OCR systems.","marker":"[Kumar et al, 2011]"},{"why":"Lists the overlapping, touching, and isolated-component phenomena that the segmentation method must handle.","marker":"[Payel Rakshit et al, 2017]"}],"fun_headline_variants":["Word and character split directly in compressed handwriting","Segment compressed scans into words and letters without decompression","90% word accuracy achieved without decompressing handwriting","Run-length encoded handwriting: direct word and char segmentation","Skip pixel unpacking: segment compressed handwriting directly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The character segmentation step assumes that adjacent letters in English handwriting overlap only in the mid-bottom region, so that overlaying the top and bottom bands reveals the cuts; if letters touch across their full height or the writing is heavily cursive, the OR pattern will not expose a separator and characters will be under-segmented.","fun_headline_variants_meta":{"raw":{"variants":["Word and character split directly in compressed handwriting","Segment compressed scans into words and letters without decompression","90% word accuracy achieved without decompressing handwriting","Run-length encoded handwriting: direct word and char segmentation","Skip pixel unpacking: segment compressed handwriting directly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000922,"raw_usage":{"total_tokens":3914,"prompt_tokens":868,"completion_tokens":3046,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2974}},"tokens_in":484,"tokens_out":3046,"duration_ms":21144,"temperature":1.0,"reasoning_tokens":2974,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:45:38.016374+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the OR-of-bands character segmentation on a set of English handwritten words whose ground truth includes many full-height touching letters, such as tightly cursive 'th' or 'll' joins; if the OR pattern shows no gap at a large share of those true boundaries, the central assumption fails. A simpler check is to measure the cascaded character accuracy on a subset of ICDAR13 marked for heavy cursive versus a subset with separated letters; the method should perform systematically worse on the cursive subset.","supporting_citations":[],"review_version":1}