{tab=Description}

¿What is The Fruit&Vegetable Image Collection?

The growing rise of new technologies and the widespread use of Internet has driven documents present a rich visual content on the web. To properly indexing these documents (images, videos, audio files, etc..) and build useful search engines, capable of meeting the information needs of users, methods are needed to accurately describe the visual and audio content.

CBIR systems ("Content Based Image Retrieval Systems") and images searchers have evolved greatly in the last decade thanks to the use of increasingly accurate descriptors in its mission to represent the image content. In fact, the synthesis of new visual descriptors is nowadays an important field of research involving institutions and companies in the search for effective solutions in image interpretation.

In order to measure the accuracy and quality of these descriptors and overall CBIR systems, is requiring specific collections of images of different types that serve as a testing ground to carry out the necessary experiments. This is the main reason that made us create The Fruit&Vegetable Image Collection.

Although over the years we have used a large number of diverse collections of images that have served to support multiple experiments, few collections that have an acceptable level of quality in terms of factors such as resolution, lighting, contrast, sharpening, no noise, etc. In particular, object-oriented collections, where the foreground power and tries to minimize the effect of noise and image background, we are not aware at present of the existence of any collection of open license with similar characteristics The Fruit&Vegetable Image Collection of quality.

As the name suggests, the purpose of the images are fruits, and vegetables, of different size, shape and color, but its main potential is its homogeneity in terms of technical features and parameters used (resolution, exposure time, aperture , zoom, ...) along the array, which greatly facilitates the evaluation of the descriptor in its task of characterizing the visual content of the image without influences of aspects accessories to said visual content.

These parameters have been selected, after doing several tests, to try to minimize the noise in the images either in the context of the image by differences in lighting or object sobreiluminación lossy. Both cases can be seen in the following image:

As seen to the left and right of the image is less bright areas while there sobreiluminadas object areas.

Although identical characteristics, each photograph is taken from a distance or at a different angle depending on their size.

Size	Distances
Small	48 cm	50 cm	55 cm	-
Normal	48 cm	50 cm	65 cm	70 cm
Big	-	60 cm	65 cm	70 cm

All these features make it excellent collection for testing for which it has been designed.

The quality of the images has been achieved thanks to the material used for this project as it is provided in a high quality SLR camera and a small studio with spotlights that has facilitated the photographs. As evidence of this can be indicative that, for example, the resolution of the images is of 5616 x 3744. The quality is also obtained product quality testing to photographs made after completion controlling noise generated and repeating as necessary.

As mentioned above all images are within the field of his Cabera Exif Copyright license has been considered optimal, which in this case has been the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.

It has also added another header field (ImageDescription) some HTML code that contains important each photograph about the texture, color, shape… A generic example of the code is as follows:

 

As seen in the image the text is in blue corresponds to the XML tags, text in black represents the possible values that can have these labels and green are the explanatory comments of its meaning.

This code gathers information about:

• Image ID
• Object ID
• Object colour
• Object texture
• Spatial attributes
• View

For identifying the image, identification, size and the view of the object have been made arbitrarily because being more specific project not found any specific ontology to define.

In contrast to the case of color, texture and shape of the object labels and their information is viewed in different ontologies:

Ontology textures	Modified
	It has adapted to the characteristics of the textures of the reasons for the images.
Color Ontology	Model HSV
Ontology forms	Modified
	You select the volume of ontology and added some form of reason itself.

 

In conclusion, the process of generating images can be summarized by the following chart:

For each image	For all images of an object

 

{tab=Gallery}

The photographic collection was performed on a total of 100 different varieties of fruits and vegetables, which because of its wide variety of sizes, shapes, colors and textures are suitable to test descriptors, additionally providing a certain appeal to images.

{tab-galeria=Collection}

The total size of the collection is of 1098 images. The collection can be downloaded in compressed format JPG and accurately jobs are also available in RAW format. Additionally it provides a file with all miniatures.

As explained in the introduction the photographed objects have a controlled size, particularly objects we selected 19 small (cherry), 67 medium-sized objects (orange) and 14 large-sized objects (watermelon).

 

{tab-galeria=Example}

All objects have been photographed from a front position and rotated 45 degrees to either the right or left, and for each position four (medium-sized objects) or three doses (objects small and large) at different distances.