Automatic defect classification without sampling and feature selection

A method for defect classification in a semiconductor process, comprising:

• extracting, using a deep-architecture neural network, one or more features from a defect image from the semiconductor process;
• injecting one or more features learned from local descriptors at one or more higher layers of the deep-architecture neural network, wherein the features learned from local descriptors are determined by:
• extracting, using a processor, a plurality of local descriptors at each pixel of each of a plurality of defect images, wherein each of the local descriptors is a defect classifier of the defect images, and wherein each of the local descriptors is external to the deep architecture neural network;
• and
• generating, using the processor, the one or more features learned from local descriptors based on the extracted local descriptors;
• and
• classifying, using the deep-architecture neural network, the defect image based on the extracted one or more features.

A system for deriving features, comprising:

• an electronic storage device;
• a feature library stored on the storage device;
• a deep-architecture neural network in electronic communication with the storage device, the neural network having one or more convolutional layers, each convolutional layer of the one or more convolutional layers followed by a subsampling layer, the neural network configured to:
• derive a feature from one or more defect images of a semiconductor wafer, wherein the feature is for classifying a defect of the defect images;
• encapsulate the feature with a set of calculations used to determine the feature;
• add the encapsulated feature to the feature library of the storage device;
• and
• inject one or more features learned from local descriptors at the one or more higher convolution layers, wherein the features learned from local descriptors are determined by:
• extracting a plurality of local descriptors at each pixel of each of a plurality of defect images, wherein each of the local descriptors is a defect classifier of the defect images, and wherein each of the local descriptors is external to the deep architecture neural network;
• and
• generating the one or more features learned from local descriptors based on the extracted local descriptors.