Choose Your Tool

The Deep Learning inspection tools are based on deep neural networks. By opposition to traditional machine vision techniques, the deep learning tools do not require an explicit model of what to recognize and/or segment inside an image. Instead, they learn this model from a set of example images. Thus, the deep learning tools can solve machine vision problems where an explicit model is too complex to build.

There are various criteria to consider when selecting a Deep Learning tool for your application, such as the type of information you want to extract from your images, the inference speed, the time you have to annotate ground truth, etc.

Comparison of the Deep Learning tools

	EasyClassify	EasySegment Supervised	EasySegment Unsupervised	EasyLocate BBox	EasyLocate Interest Point
Inference speed	++++	+	++	+++	+++
Prediction	Unique image label	Labeled segmentation map	OK/NOK segmentation map	Labeled bounding box	Labeled points
Additional capability	Heatmap		Capacity to detect unknown defects[1]	Distingish overlapping objects, counting[2]	Counting[2]
Ground truth annotation	Image label	Per-pixel label	OK/NOK image label	Labeled bounding boxes	Labeled points

---

[1]

EasySegment Unsupervised trains only on images of a good product and identifies images of defective product by how much they deviate from images of good products. The advantage is that the tool can detect and segment defects that are not in your dataset and that you did not expect. The drawback is that the type of defects that the tool can detect is more limited compared to the other tools.

[2] (1,2)

EasyLocate works well with partially occluded objects or with objects overlapping each other. You can use the EasyLocate tools to do counting applications.