Fishial.AI is a project initially sponsored by "The Wye Foundation" with the goal of making highly accurate fish identification "Fishial Recognition" possible. Our goal is to create an open source model that uses computer vision to successfully identify fish by species through images submitted to the model.
Currently, large databases of correctly labeled fish images for machine learning does not exist so we decided on building on own to achieve the level of accuracy we desire in the algorithm. After unsuccessfully seeking out a large dataset of labeled fish images we recognize the need and value of such within the AI community. We believe by creating a large ML enabled fish image collection will drive the AI community in making Fishial Recognition a reality!
Image-net, COCO and Google are image databases within the AI community; however, they lack large sets of label fish images. Image-net has just over 1300 labeled fish images, COCO does not have any and Google has a few thousand labeled fish images. Unfortunately, not even combined could these sources provide enough images to train a AI model robust enough to identify fish species worldwide. By creating a crowd sourcing web-based platform the Wye Foundation has uniquely positioned themselves to solve this issue, by creating a place where fish images can be collected and labeled.
In order to achieve this, we are building the world’s largest fish image database, use training images that have been identified with fish species and have been tagged with fish species attributes unique to the species. The training images will also have polygonal style bounding boxes only around the fish in the image, with anticipation that this will reduce background noise for the model and allow for a quicker and more accurate training process. We anticipate that by training the model with images that have been tagged with these unique identifying attributes for each fish that we will be able to create a model that will be able to identify fish species within images with high probability. We are hoping that Fishial recognition has broader significance in helping to identify fish species that are not frequently identified and or help identify those species that have not yet been discovered.
At the moment the Wye Foundation has completed the first phase of the Fishial.AI project, to build a crowd sourcing web-based platform that allows for the collection of fish images and labels. The Wye Foundation is seeking assistance in the second phase of the project which focuses on building the worlds largest dataset of labeled fish images. In order to do this, we will need to assemble a team of knowledgeable individuals that will help collect fish images and label the images with fish species.
Individuals and teams of interested parties are be able to contribute to the Fishial.AI photo library by joining the portal. Once logged into the portal the user are be able to upload as many images as they wish, label fish species in the images, tag the unique identifying attributes and set their preferences for creative commons.
The open-source image database will allow for researchers worldwide to use the fish images in creating their own AI models, thus accelerating the development of AI in solving fisheries science related issues.