Image Recognition Vs. Object Detection
Image Recognition has become quite noteworthy as of late due to all of the attention generated by advancements in Facial Recognition. However, for the majority of businesses, it is Object Detection, as well as Special Feature detection, that is used to solve for the majority of use cases we have seen at Macrosoft AI. Our solutions allow for any combination of object definition/categorization, labeling, filtering, editing and parsing of any image.
Transfer Learning in Model Training for Object Detection
Success in the delivery of effective Object Detection solutions is always challenged by the need for very precise model training. In the majority of cases, one does not have to start from a “clean slate”, i.e. starting with a completely untrained model.
For example, a telecommunications customer recently acquired a large new territory for its service footprint. This company asked Macrosoft AI to help them identify the type of dwellings – both residential and commercial – that reside in this new market. Additionally, in scope were many specific dwelling attributes (i.e. special features) that were relevant to the sale and/or delivery of the client’s services. This had direct impact on their sales, marketing, pricing and other tactics are dependent on the type of dwelling or structure is to receive their service. It would have been ineffective from a time and cost standpoint to create a solution where we needed to “train” our model from scratch on what is a house, what is a high rise, etc. In this case, Macrosoft AI implemented a classic form of Transfer Learning; the use of a pre-trained model. We used Google Vision as a starting point, which was then refined by the use of detailed training “data” (i.e. images) to create the level of sophistication required for the solution.
Deep Learning in Model Training for Object Detection
Not surprisingly, however, there are certain use cases where the object(s), or special features of interest, have not been previously modeled to any extent, preventing the use of classic Transfer Learning as in the example above. In such cases, Macrosoft AI turns to the application of Deep Learning: the creation of a neural network of models, connected in a manner similar to that of the human brain.
Deep Learning is an expansive and highly complex AI solution methodology that is deserving of a much larger discussion than is practical here. Suffice it to say that, within the context of Image Recognition, Macrosoft AI utilizes Deep Learning when a use case requires our solution to “learn from scratch”. Such solutions require us to acquire our own training “data” (i.e. images) to utilize within our neural network. As such, Macrosoft AI serves as both a data as well as a solutions provider.