<\/p>\n\n\n\n
Offhand, it sounds nuts. Even an experienced pashmina vendor is sometimes unsure about a fabric that he or she is actually touching. How could a conventional photograph possibly be more certain?<\/em><\/p>\n\n\n\n Before pashmina can be assessed, the term has to be defined. The word may refer to the fine undercoat of a cashmere goat, yarn made from that fiber, cloth containing cashmere\u2014sometimes blended with silk or wool\u2014or simply a large shawl marketed as a \u201cpashmina.\u201d Claims made for \u201creal,\u201d \u201cpure\u201d and \u201cauthentic\u201d pashmina add further requirements involving goat breed, geographic origin, fiber diameter, staple length, hand-spinning, handloom production or Kashmiri tradition. These are not the same claim, and no single test necessarily answers all of them. Yet buyers and dealers rarely have access to laboratory analysis capable of answering even the narrower question: what fibers does this fabric contain?<\/p>\n\n\n\n A striking alternative to the laboratory is proposed in \u201cPashmina Authentication on Imagery Data Using Deep Learning<\/a>,\u201d by Muzafar Rasool Bhat, Assif Assad, Ab Naffi Ahanger, Shabana Nargis Rasool and Abdul Basit Ahanger. Published online in 2023 and in print in 2024, the article appeared in the Open Forum section of AI & Society: Knowledge, Culture and Communication<\/em>. This is an established, refereed, interdisciplinary journal published by Springer Nature and indexed by major scholarly databases. It is a recognized scholarly venue, although neither the journal nor its Open Forum is devoted specifically to textile science. The article has so far attracted modest visible attention: as of October 2026, the publisher reported 529 accesses, and the citation indexes I found recorded one citation.<\/p>\n\n\n\n The five authors were affiliated with the departments of Computer Science and Computer Science and Engineering at the Islamic University of Science and Technology in Awantipora, Jammu and Kashmir. Their published work is principally in artificial intelligence and image classification rather than textile science, although this was not their first attempt to apply those methods to pashmina: members of the group had already filed a patent application for a system to detect and classify pashmina shawls. The research was funded by the Jammu and Kashmir State Council for Science and Technology. Their AI model, PashmNet, reportedly classified photographs as \u201cPashmina\u201d or \u201cnon-Pashmina\u201d with 97.4 percent accuracy. They envisage a non-specialist being able to photograph a fabric with a conventional camera or smartphone and receiving a judgment of \u201creal pashmina\u201d or \u201cfake pashmina\u201d\u2014a cheap, rapid and widely distributable substitute for scarce expertise and expensive laboratory testing.<\/p>\n\n\n\n That is the claim assessed here. I will first consider what the authors mean by \u201cpashmina\u201d and how the original fabrics were authenticated; then examine the photographs, data preparation, reported accuracy and possibility that related images entered both training and testing; and finally ask whether PashmNet has been independently tested, developed into a usable tool or shown to work on unfamiliar shawls photographed under ordinary market conditions. My perspective is that of a long-time pashmina dealer and student of the trade, not a textile laboratory scientist or machine-learning specialist. I have not inspected the fabrics, dataset or software. This review therefore cannot independently test PashmNet; it can assess whether, and how far, the published evidence supports the practical claims made for it.<\/p>\n\n\n\n The available scientific methods are not imaginary. Cashmere and other animal fibers may be examined microscopically; more specialized methods include spectroscopy and DNA analysis. Kashmir has also developed certification and labeling systems intended to protect its pashmina industry. But these methods require equipment, expertise, access to suitable facilities and, in some cases, removal of material from the product. They are not ordinarily available to someone considering a shawl in a shop, a dealer evaluating a shipment or a consumer questioning an online purchase.<\/p>\n\n\n\n In practice, buyers rely heavily on labels, reputation, touch and the assurances of sellers. Popular advice adds home tests involving burning, shine, static electricity, itching or whether a shawl will pass through a ring. Such tests range from inconclusive to destructive, and several test properties other than fiber content. A dependable photographic method would therefore be genuinely important. It could screen products instantly, without damaging them, and make at least one form of expert judgment widely available at negligible marginal cost.<\/p>\n\n\n\n The paper makes this practical promise explicitly. The authors write that a person could photograph a fabric and have the system recognize \u201creal Pashmina\u201d or \u201cfake Pashmina.\u201d They suggest that the resulting confidence would protect buyers, reward sellers of genuine products and strengthen the pashmina economy from traders and artisans back to goat herders. They also say that large quantities of images could be processed rapidly and cheaply, making the method scalable for commercial use.<\/p>\n\n\n\n Those advantages depend, however, on what the two classifications actually mean.<\/p>\n\n\n\n The paper begins broadly. It calls pashmina a special wool obtained from cashmere goats, and it notes that cashmere is produced by several goat breeds in China, Mongolia, Russia, India, Turkey and Pakistan. It then narrows the term: the Changthangi goat of Ladakh, it says, produces \u201cthe specific Cashmere fibre known as Pashmina.\u201d Elsewhere it moves among pashmina fiber, fabric, apparel and shawls. \u201cAuthentic\u201d pashmina is associated not only with fiber content but with hand production, Kashmir and a particular cultural tradition.<\/p>\n\n\n\n This terminological movement is common in the trade, but it creates a serious experimental problem. If pashmina means cashmere fiber, the relevant question is whether the cloth contains cashmere and in what proportion. If it means fiber from Changthangi goats, the question includes breed and perhaps geographic origin. If \u201cgenuine Pashmina\u201d must also be handloomed in Kashmir, manufacturing method and provenance become part of the classification. Those are different propositions requiring different evidence.<\/p>\n\n\n\n Some of the paper\u2019s background claims further blur these categories. It says that pure pashmina can be made only on handlooms because delicate fibers cannot withstand power-loom vibration without being blended with silk, merino or angora. That is an assertion about manufacturing as well as composition. The discussion of shine, static and itching similarly treats a collection of sensory or behavioral cues as indicators of adulteration. Even if one of these cues correlated with the authors\u2019 samples, it would not necessarily identify the fiber responsible.<\/p>\n\n\n\n The critical issue is therefore not philosophical hairsplitting. An image classifier can only reproduce the categories assigned to its training photographs. Before interpreting a PashmNet result, we need to know what qualified a source fabric for the label \u201creal Pashmina\u201d and what disqualified one as \u201cnon-Pashmina.\u201d<\/p>\n\n\n\n The researchers began with 330 fabric photographs: 130 labeled real Pashmina and 200 non-Pashmina. They broke the photographs into patches and manually selected patches judged to represent substantial portions of the fabric. This produced 4,297 images. They also augmented the data by creating shifted, flipped and zoomed versions, then report dividing the dataset into training, testing and validation sets in a 65:20:15 ratio.<\/p>\n\n\n\n Four models were compared: a basic convolutional neural network and models based on ResNet50, VGG16 and InceptionV3. ResNet performed best, with reported accuracy of 97.4 percent. VGG16 reached 95.8 percent, InceptionV3 85.7 percent and the basic model 68 percent. The comparison provides evidence that transfer learning\u2014and ResNet in particular\u2014classified this dataset more successfully than the alternatives tested.<\/p>\n\n\n\n That is a useful result. It is narrower than showing that a photograph can establish the fiber content of a shawl.<\/p>\n\n\n\n A photograph may reveal weave, color, texture, embroidery and finishing. It may also contain unintended clues arising from lighting, camera, background, cropping or the way different samples were handled. The experiment does not establish which features drove the classification, whether those features are unique to cashmere, or whether they would remain reliable when the same material was woven and finished differently.<\/p>\n\n\n\n This is the paper\u2019s most consequential omission. The methods section does not say where the 330 original fabrics came from, how many distinct physical pieces they represent, who classified them, what tests or documentary evidence established their composition, or how uncertain cases were treated.<\/p>\n\n\n\n The model cannot discover the meaning of pashmina independently. It learns statistical differences between images carrying labels supplied by the researchers. If the \u201creal\u201d set consisted chiefly of traditional Kashmiri shawls and the other set of visibly different fabrics, the model might become highly proficient at recognizing differences in weave, finish, color or photographic treatment. It could reproduce the original labels with impressive accuracy without detecting cashmere fiber.<\/p>\n\n\n\n Ground truth is therefore not a preliminary detail; it is the foundation of the experiment. Ideally, every source fabric would have independently verified fiber composition, with breed or provenance documented separately rather than folded into a single binary label. The test set would include cashmere from different breeds and origins, blended cashmere fabrics, non-cashmere fabrics designed to resemble pashmina, and pashmina made using different colors, weaves and finishing methods.<\/p>\n\n\n\n The authors acknowledge one portion of this limitation. Detecting blends and estimating the amount of impurity are described as future work. But that admission sharply restricts the present claim. Much of the real market is not divided into 100-percent pashmina and zero-percent pashmina. Cashmere-silk blends are legitimate products when accurately labeled, and fraudulent products may contain some cashmere while claiming much more. A system unable to identify blends or estimate composition cannot yet answer the question most buyers need answered.<\/p>\n\n\n\n The paper\u2019s 4,297 images came from 330 original photographs. The additional images were patches and augmented versions, not thousands of independently sourced fabrics. This distinction matters because closely related images are not independent evidence.<\/p>\n\n\n\n The paper describes patching and augmentation before describing the final division into training, testing and validation sets. It does not state whether every patch and altered version derived from one original photograph\u2014or from one physical fabric\u2014was kept within a single division. If related patches entered both training and testing, the model may have been tested on another view of material it had effectively already seen.<\/p>\n\n\n\n This is a risk raised by the reported workflow, not proof that leakage occurred. It could be resolved by dividing the original physical fabrics into training, validation and test groups first, and only then producing patches and augmentations within each group. The strongest evaluation would go further: an independent group would test the completed model on an entirely new collection assembled and authenticated without the model developers\u2019 participation.<\/p>\n\n\n\n The published numbers also contain an unresolved discrepancy. A 20-percent test set drawn from 4,297 images would contain approximately 859 images. The results section instead says that the test set contained 2,000 images and that this represented 20 percent of the entire dataset. Two thousand is about 46.5 percent of 4,297. This might be a typographical error or evidence of an additional, unreported stage of dataset expansion. Either way, the account cannot be reconciled as published.<\/p>\n\n\n\n The researchers used a method called Grad-CAM to visualize which regions of an image most influenced the model. They report that the model activated across the fabric rather than focusing on particular embroidery, and interpret this as confirmation that the network behaved as expected.<\/p>\n\n\n\n The visualization is useful, but its evidentiary reach is limited. Showing that the model looked broadly at the cloth does not reveal whether it responded to cashmere fibers, weave density, fuzziness, color, finish or some photographic artifact distributed across the image. A heat map locates influential pixels; it does not identify the physical property represented by them.<\/p>\n\n\n\n Controlled comparison would be needed to separate these possibilities. The same fabrics could be photographed with different cameras, backgrounds and lighting; different fibers could be woven and finished to look alike; cashmere from different sources could be presented in dissimilar forms; and every classification could be compared with independently established fiber composition. The paper does not report such tests.<\/p>\n\n\n\n The dataset is available from the corresponding author \u201con reasonable request,\u201d but it is not deposited in an open repository. The code and trained model are not linked in the paper. That does not invalidate the experiment, but it makes independent inspection, replication and improvement more difficult.<\/p>\n\n\n\n The authors\u2019 computing background is relevant but should not be treated as a disqualification. Image classification is precisely the expertise needed to construct PashmNet. Their broader records include applications of deep learning to crops, medical imagery, rivers and other classification problems. The question is whether complementary textile expertise and reliable sample authentication were incorporated into this particular project. The paper does not identify a textile scientist, testing laboratory or pashmina-certification body as a collaborator.<\/p>\n\n\n\n Nor was the project merely invented for a journal article. IUST records show a 2022 patent application titled \u201cSystem and Method for Detecting and Classifying Pashmina Shawl,\u201d involving several members of the group. Four of the article\u2019s authors were also listed on a later 2022 application concerning identification of handwoven and machine-woven fabrics. An earlier software registration concerned AI-based identification and classification of pashmina artwork. These records show a sustained attempt to develop commercial or practical applications around visual textile classification.<\/p>\n\n\n\n What they do not yet show is successful deployment. Searches through October 2026 found the published paper and patent records, but no publicly available PashmNet application, open model, independent field trial or follow-up study validating the method on new pashmina products. The paper itself has circulated within the scholarly indexing system but has attracted little visible academic or public attention. The project is therefore neither frivolous nor, at present, an established authentication service. It remains a promising and insufficiently validated proposal.<\/p>\n\n\n\n PashmNet demonstrates that a transfer-learning model can distinguish two labeled groups of fabric images in the dataset assembled by its developers. That is an interesting proof of concept. The 97.4-percent figure, however, describes performance within that particular experiment. It is not the probability that a shawl photographed by a customer is genuine.<\/p>\n\n\n\n Before PashmNet could perform the role proposed for it, the researchers would need to define exactly what it authenticates; document how every source fabric was classified; separate data by physical sample before creating patches or augmentations; resolve the test-set discrepancy; and evaluate the completed model on independently sourced fabrics photographed under ordinary conditions. A useful consumer system would also need to address blends and report uncertainty rather than forcing every fabric into an unqualified real-or-fake judgment.<\/p>\n\n\n\n The underlying ambition deserves attention. Laboratory testing is effectively absent from most pashmina transactions, while labels, touch and inherited folklore provide uncertain protection. A dependable, accessible screening tool could improve the market considerably. PashmNet suggests that ordinary photographs may contain useful information. The published evidence does not yet show that they contain enough information to verify pashmina content.<\/p>\n\n\n\n Bhat, Muzafar Rasool, Assif Assad, Ab Naffi Ahanger, Shabana Nargis Rasool, and Abdul Basit Ahanger. 2024. \u201cPashmina Authentication on Imagery Data Using Deep Learning<\/a>.\u201d AI & Society<\/em> 39: 2297\u20132305. Published online May 4, 2023.<\/p>\n\n\n\n Federal Trade Commission. \u201cCachet of Cashmere: Complying with the Wool Products Labeling Act<\/a>.\u201d<\/p>\n\n\n\n Islamic University of Science and Technology. Faculty profiles and patent records<\/a>.<\/p>\n\n\n\nWhy a photographic test would matter<\/h2>\n\n\n\n
What does \u201cPashmina\u201d mean here?<\/h2>\n\n\n\n
What the model learned<\/h2>\n\n\n\n
Who authenticated the authentic samples?<\/h2>\n\n\n\n
Four thousand images are not four thousand fabrics<\/h2>\n\n\n\n
A heat map is not a fiber test<\/h2>\n\n\n\n
The researchers\u2014and what followed<\/h2>\n\n\n\n
A useful beginning, not yet an authenticator<\/h2>\n\n\n\n
References and sources<\/h2>\n\n\n\n