As an ongoing development for automation continues, new technology is being developed everyday with the purpose of lessening human interference. This being said, one issue that can be resolved is the misidentification of snake species. As one can predict, many species look visually similar to the untrained eye, so developing software to help non-professionals can bring awareness to the various species. The development of deep learning networks is vital to the improvement of image segmentation when data is being drawn for quantitative purposes. A variety of pre-made algorithms and applications for segmentation have been previously discussed in literature, but many of which have not been implemented for the purpose of image analysis on snakes. Such analysis can lead to more development in technologies used for snake identification between venomous versus non-venomous species.
As of today, few resources have been published on the interest of scale segmentation in snakes. With that being said, more automatic image analysis to detect species of snakes hasn’t been made available for public use. In the machine learning community, little focus has been brought to the interest of image segmentation within snake species as large datasets are not currently available [1]. To train models for recognition, large datasets are needed for the model to learn with more information. There are, however, attempts to take on these challenges such as using deep learning networks and ImageNet pre-trained checkpoints [2]. One of the outcomes from this attempt is the comparative analysis between different types of machine learning tactics for snake identification, which resulted in the best technique using a large visual database rather than a smaller one.
I’ll use machine learning and implement it for native Nebraska snake identification. This tactic will involve building a neural network with prelabeled images to train datasets for the computer to automatically differentiate key components in a diverse panel of images. It will require human intervention to determine the set of features that will help the software understand the differences between images. With a high accuracy rating in identification from the initial set of images, further investigation can occur that can improve the reading of native Nebraska species.
This strategy of identification can be implemented in other fields such as agriculture, medical, conservation, etc. Specifically in regards to this project, the goal for identification of native Nebraska snakes can be expanded to the range of native Midwestern or even North American species. As the range of what species could be identified stretches, more datasets will need to be labeled for each individual species. With success, similar structures compared to snake scales (i.e. oversimplified cells in plants or tassels on maize) can be quantified automatically without manual labor. Individuals with little to no experience in snake identification can be educated with an image. With failure, it will let me know that machine learning isn’t the best tactic for identification and other strategies can be tested such as computer vision.