Welcome back to my occasional image classification series.

The last time I raised the topic of image classification I discussed the basic attributes of images. This time I want to focus on the thornier issue of the content, or concepts, depicted in them.

There is a danger of treating an image like a piece of text and classifying its attributes: Who created it? When? What techniques were used? Then writing a title or caption and leaving it at that. Sometimes little more need be done to a document than record this kind of information, especially with free text searching, but lots more needs to be done to most images.

Image findability

Image findability is the process of using search and browse to access the images required. A major aspect of image findability relates to the things depicted in them. Image users often search for images based on the generic things in them and also the proper names of these things. Classifying images based on depicted content means considering anything and everything that is and can be depicted in an image. When considering this I like to focus my efforts on understanding the images I'm dealing with, the users who are trying to find and work with the images, and the ways in which these people need to search and browse for the images they need. After an assessment of these areas I then tailor my approach.

Broadly speaking people searching for depicted content are looking for a number of types:

• Places: cities, towns, villages, streets...
• Built works: parks, skyscrapers, cottages, walls, doors, windows...
• Topography: mountains, valleys...
• Groups and organisations: air forces, choirs, police departments...
• People: roles, occupations, ethnicity and nationality: mothers, doctors, Caucasians, French, Germans...
• Actions, activities and events: running, writing, laughing, smiling, birthdays, parties, book signings, meetings...
• Objects: a myriad of items...
• Animals and plants: common and scientific names...
• Anatomy and attributes of people, animals and plants: arms, legs, adults, leaves, trunks, paws, tails...
• Depicted text shown in images - often signs or writing shown in images...

Many of these generic types can also have proper named instances:

• Proper names of people, places, buildings, topography, organisations, animals etc

When dealing with depicted content I've found some of the biggest issues to be:

• Identification - knowing what is in an image
• Focus and specificity - knowing what to include and what to exclude
• Consistency - applying the same term in the same way for the same depicted content

Identification - knowing what is in an image

Depicted content is a relatively black and white area - a dog is depicted so a dog is tagged. However, it might sound a little weird, but working out what is actually in an image can be a lot harder than you think.

Take a look at the image "Do You Know What This Is?" by Sister72

This depicted content is fairly simple to see, but understanding what you're looking at is not that easy. Even if you know roughly what you're looking at, do you know what it's actually called?

One tip is to group similar images together when you're classifying them. Also, always start by assembling as much information as possible before you begin to classify images. It is especially important to gather together the information you have from the creator or custodians of the images.

Also important, when you have the luxury, is to get the image creator to add key metadata about the image at the point of creation, or soon after.

Focus and specificity

Knowing what to include and what to exclude, what to mention and what to ignore, is also much harder than it sounds.

Firstly, some image users will want a piece of depicted content tagged whenever it appears in an image, others will only want it tagged when the image shows a very good representation of that content, and of course many people will want something in between the two extremes.

Different users have different requirements. You need to understand the domain in which you're working and see the classification of depicted image content as supporting the needs of your users.

For example, Would you tag everything in this 'Messy Room' image?

What would you miss out and why?

Looking at the image of "Mountain Goats", from Thorne Enterprises

Would you tag this with goats as well as mountains? Would this be helpful?

Let's look at four images depicting windows:

'Window to the World'?,

'Portuguese Window'?, '

What Light Through Yonder Window Breaks'?

and

'Window'.

Looking at these, it soon becomes clear that even deciding to apply a simple term like 'Windows' is not always easy.

Would you apply 'Windows' to the image of the cat looking out of the window? Is a window actually depicted in that image? If the image wasn't tagged with 'Windows' how else would anyone find an image of a cat looking out of a window?

The other three images show windows as parts of buildings. but is a building always depicted? Deciding when to apply a building type or the name of a building can be hard. Should you do this every time a part of a building is shown? Only when the whole building is shown? When enough of the building is visible? Or when a section of the building that to most people would represent the build is visible? For example, what part of the Empire State Building would you consider to depict that building? Rarely does anyone see it all - how much is enough? Would you treat the images of windows in a similar way and classify them all with a building type of 'Houses', or would you ignore the structure and focus on the parts - the window, the roof?

Consistency

Achieving consistent application of terms to images revolves partly around clear term definitions, well defined application rules and guidelines, and a robust quality assurance process.

Term definitions are very important. Defining the meaning of a term, and ensuring the people choosing which term to assign understand that meaning, can be crucial to term application. For example, creating a term such as 'Bow' without defining its meaning is not going to make it easy to apply.

Application rules that are well considered, thorough and clear are also very useful. Even a simple concept often needs some form of guidance linked to it. I remember a while ago needing two terms, 'Indoors' and 'Outdoors' to allow users to find images of people who were outside and inside - a simple concept you might think, one that people often need, and one that's easy to apply - who'd need guidelines for that? However, it soon became clear that guidelines were needed after I received a series of interesting questions: Is being on a train indoors? Should studio shots always be considered indoors? Does every shot of a person have to have indoors or outdoors assigned to it? If not, when should this term be used and when not? Is this a focus issue? If so, how much of a location needs to be seen before Indoors or Outdoors is used. A clear set of application guidelines followed an interesting meeting!

Strong quality assurance processes are very valuable. People make mistakes and images generate interesting issues. Appointing staff to review a percentage of classification work based on clear guidelines, and then sharing findings with the people who assigned the terms to the images, is an important way of assessing how well the image classification is progressing and keeping a classification team synchronised.

Today I’ve talked a lot about content depicted in images, next time I’ll focus on abstract concepts which are related to an images ‘aboutness’.

We are very excited to announce one of the many new features that will be available with our next release of Synaptica Version 7.1 (available mid-July). This new feature will provide a mechanism that will doubtless save tens if not hundreds of hours for those that currently do manual "mappings" between two different vocabularies.

The new Term AutoMatch tool will allow for the comparison of terms between two distinct Object Classes (vocabularies) and automatically perform a "mapping" between terms where there is a match based on whole or part of the term descriptions. The feature first allows for the selection of the two Object Classes and then for the matching method one would like to apply.

As you will see in detail in the above slidedeck, there are five matching levels including exact match, keywords match, a soundex match (based on phonetics) and finally a single or multiple word smart matches. One may then select the relationship type that should be applied between the terms as well as which types of matches should be displayed.

• Display all matches
• Display matches except where match already exists
• Display matches with no existing match from first to second object class

Next, one may use the built-in review tool to individually remove invalid matches and immediately submit the matches as new relationships in Synaptica. Optionally, the match results can be downloaded, adjusted within a spreadsheet and then re-uploaded to Synaptica to create new relationships. For the online version each term name is displayed as a link so one may enter the Item Summary for any term and edit the term directly. Each term will display an "Existing Relationship Count" that shows the number of relationships from that term to terms in the other object class.

 This tool also uses synonyms to allows for matching between “preferred” and “non-preferred” terms, finding matches using synonyms that might otherwise be missed. A tips screen is available to assist with all of the features and the overall use of this tool.

 We have already had tremendous response from customers that have seen an advance release of this new feature. We expect that it will be a tremendous benefit to those that perform this type of task on a routine basis and look forward to delivering it to customers.

 Please contact us to find out more about Synaptica AutoMatch or for a demonstration and of course stay tuned for more announcements on Synaptica's upcoming features!

I was very interested to see Google experimenting with visual similarity in still images, what I usually call Content Based Image Retrieval or CBIR.

Google Labs have just launched an image search function based on visual similarity - Google Similar Images. This new offering allows searchers to start with an initial image and then find other images that look like their example picture.

I've been reviewing these type of systems on and off since the early '90s. They've always offered much, but I never saw any evidence that the delivery matched the hype.

I've always found that using pictures instead of text to find images works best on simple 2d images: carpet patterns, trademarks, simple shapes, colours and textures. Finding objects in images was always a struggle, and looking for abstract concepts: fear, excitement, gloom, isolation, solitude.. was never been more than a vague possibility. Over the years a lot of work has been done in this area, and the search results I've seen have started to improve, but this technology is still young, and in my personal opinion still rarely delivers what most users want, need and expect.

Looking at Google Similar Images, I wonder how much of the back-end is pure content based image retrieval (CBIR), how much is using metadata in some way, and how the two are interacting? One thing that appears to be helping to often show a tight first page of results, is simply pulling the same image from different sites. I also noticed that the 'similar images' option is not available for all images - which makes me wonder why? Have some images been processed in ways that others haven't?

Diving right into the experience, I entered a query for a place in the UK and didn't see any image results with the 'Similar Images' option. I wonder whether this is to do with the presence of the results on UK websites?

I persevered, and found some interesting images and got some interesting results.

I started with a fairly standard image of a beach scene, always a favourite with testers. As you can see I got a pretty good first screen back. However, the 5th and 6th image on the top row show no sea or beach, neither do the first three images on the second row.

I moved on to an image of what looks like equipment at the top of a pole.

The results were much more mixed: studio shots of objects, fighting people, trucks etc. No images were returned that I would consider similar to the example picture.

Interesting results came from a similarity query on a clock face.  A couple of the first results hit the mark, then the results set degenerated into image similarity based more on the colour and the black background than anything else.

My last attempt, before morning coffee called, was an image of a country road. I was hoping that the clear roadway might produce a pretty precise results set. However, I was a little disappointed by what I saw.

The first results page only produced one vague road on the bottom row, with most of the similarity seemingly related to colours instead of objects.

From my less than scientific dip into this Google Labs offering, it looks like the highlighted images on the Google Similar Images home page produce good results - better results than I've seen other systems come up with. Many other image queries are sure to also produce results which may well impress. However, many of the results I saw did not match the initial level of accuracy I saw from the highlighted home page pictures.

I don't want to be picky, this is still a prototype after all, and well done to Google for introducing a wider audience to this type of image search. Hopefully, after more work, the results will increasingly make more sense to people, the access points offered to depicted content and conceptual aboutness will improve and more images will be more findable for more people.

Until that time, visual search without text will help with image findability, but text, metadata, and controlled vocabulary applied to images by people is for me still king, and will continue to offer the widest and deepest access to images for a long time to come.

Ian

I recently attended a seminar on the 10 Steps to Business Efficiency with Content, Collaboration and Process given by the good people at AIIM (http://aiim.org) all about ECM strategies and best practices. This was a free seminar, well organized and well attended by a broad spectrum of representatives from all types of organizations, large and small, new and old industries. The topics of discussion too ranged from the most effective way to digitize archival assets; to applications to better allow for federated search across various data repositories; and then there was certainly a lot of discussion around what has become the most ubiquitous of ECM type applications, Microsoft SharePoint.

There were of course the usual quotes and statistics from AIIM, Forrester and Gartner regarding information proliferation and management today: The amount of data being produced is doubling every 18 months; 80% of this data is unstructured and 90% of that is entirely un-managed.

An interesting quote that I will paraphrase here was attributed to Thomas Washington , "The pursuit of knowledge in an age of information overload is less about the process of acquisition than it is about a proficiency of tossing things out." And regarding the storage of all of this information another interesting fact was thrown out: while 1 GB of storage may now cost an average of 20 cents, it costs $3,500 to review that same 1 GB of data and start to make sense of it in the context of your business. (AIIM)

As I listened to the various presentations and vendors I was struck by one thing: none seemed to offer a unified solution for using taxonomy more effectively to structure, classify and categorize the content that was going into these vast data repositories. Certainly it was agreed that there was value to such a process, but it is something that many organizations have still not recognized as absolutely necessary to fundamentally improve the tagging, organization and discovery of information within these huge libraries of data, documents, and other media.

It is our opinion that the integrated use of taxonomy applied to ECM applications, as well as across the rest of the enterprise, using a centralized and standardized set of vocabularies for navigation, search, discovery, meta-tagging and many other applications is a necessity in moving towards a unified means of data normalization and discoverability. To achieve this we offer services to get companies started as well as tools like Synaptica with out-of-the-box integrations to tools like SharePoint, but also more generic means of integrating with external applications via simple APIs and Web Services.

As the proliferation of data only increases over time and the means of digitizing archival records or utilizing native electronic formats becomes more efficient, storage becomes less a matter of cost and more a matter of management. The efficient means of identifying, tagging, categorizing and sorting information will be key to the effective operation of any organization.

A couple months back, my colleague also wrote up the 10 Rules of Successful ECM Implementation after attending an AAIM seminar that we have found quite useful in talking to business and technology owners about content access strategies.

We see many of our customers at the forefront of addressing these issues and working with them, we continue to work towards providing better and easier ways for data managers and end users alike to find what they are looking for. We look forward to sharing some of these use cases as well as hear from you on your successes and struggles!

Image| Flickr | ul Marqa

Our colleagues at Factiva have decided to start their own blog, Let's Talk Knowledge - congrats gang! Might I humbly say that they were inspired by Synaptica Central, and thank YOU, our readers, for helping make this enough of a success as to inspire others here at Dow Jones!

Ken kicks off this new endeavor with some interesting thoughts on "Information Overload." The problem, he posits, is not that we have too much information, but have not invested enough in the tools to manage and analyze the information. IT investments have focused on creating, storing and distributing information (and, I would argue, retrieving it), but NOT in how we analyze or synthesize it. As Daniela and I have argued here on Synaptica Central in posts on data visualization and our Semantic Webinar, that is truly an area that needs work!

Let's be clear though: it's not about the latest widgets! It's not about the pretty, sparkling thing that just flashed in the browser window! It's about data that can be used and re-used in any form - be it sparkly, conservative, mobile or consumed by machines. Putting time and effort into crafting the models for this data can reduce confusion, reduce time spent towards rules-writing or custom query building, and allow a greater diversity in a product portfolio from the same data set: not just in delivery channels, but in customer focused channels as well. Employees are more productive, prouder of their work and motivated to innovate. Customers get what they need - they reduce information overload as well as relevancy overload - and are much happier and more loyal for it.

We're working hard here at Dow Jones to integrate and evolve our data models to full take advantage of the semantic web. Taxonomies and thesauri are becoming ontologies; lightweight ontologies are being transformed to take advantage of the full power of RDF, OWL and SKOS. Welcome to the blogosphere Ken, thanks for your thoughts, and we look forward to continuing to innovate with you and your team!