This AI Engineering Computer Vision Service is intended to be a complete insightful source of detection through of a Convolutional Neural Network Classificator. Furthermore, it has the purpose of bringing a business/educational differential advantage to the Bug's Life application.
This model recognizes and automatically identify if an insect exists or not in a given image.
The first objective was reached in which a deep learning model was trained successfully, The model was deployed on Render through Docker, Flask, Starlette.
This model pretends to recognize species, families and orders of the insect world.
Dataset from adults insects and caterpillar from butterflies; specific angle of view and distance; insects without persons, laboratories and homes in the photo; photos in the forests on a daylight.
Total: 6000 Images
8 classes of Convolutional Neural Network is:
Bee, Ant, Cockroach, Dragonfly, Butterfly, Fly, Bettle, Others (1st Version)
Total Augmented Images: 25600 Images
Make some annotations in IBM Cloud using about 100 images could help me to automate the process in other thousands insects images. I've made a Machine Learning model to learn initial annotations and predict new ones.
Using methods like Cutoff and Mixup directly on training dataset, to insert a dropout in initial layer of the Neural Network A[0]. Also other techniques: (Rotate, Brightness, RandomCrop, Jitter)
Constructing a NN with correct BatchSize(32,64) hidden layers(ConvLayers), activations (ReLU), droupouts and normalizations (imagenet_stats) to achieve the best result. I'd rather use 2 kinds of Frameworks to construct the NN (Keras, Fastai) to ensure the goal of model.
I made a research to visualize better ways to achieve the best accuracy in the model. Applying techniques like Mask R-CNN and Active contour model, I'm researching how to use instance and semantic segmentations in specific classes of insects.
1- Train the Deep Learning Model (First Version)
2- Hyper-parameter Tuning
3- Evaluate with Real tests
4- Deploy