MUMBAI, India, May 1 -- Intellectual Property India has published a patent application (202641050199 A) filed by Indian Institute Of Technology, Hyderabad, Telangana, on April 20, for 'complementary metal-oxide-semiconductor (cmos) image sensor with convolutional neural network (cnn) implementation on analog domain.'
Inventor(s) include Jainwal, Kapil; and Kumar, Aryan.
The application for the patent was published on May 1, under issue no. 18/2026.
According to the abstract released by the Intellectual Property India: "The present disclosure provides a method (900) for implementing a convolutional neural network (CNN) on an analog domain, comprising selecting (901) a complementary metal-oxide-semiconductor (CMOS) pixel window from a CMOS pixel array, and performing (902) a convolution operation on the selected CMOS pixel array using a first switch-capacitor based adder. Further, a rectified linear unit (ReLU) operation is performed (903) on an output of the first switch-capacitor based adder, wherein the ReLU operation and the convolution operation form an input layer. Further, an output of the input layer is transferred (904) to a hidden layer, wherein the hidden layer comprises the steps of employing (905) a max-pooling operation, wherein the max-pooling operation selects a maximum value from each set of a plurality of input values, and scaling (906) the pooled maximum values by a switch-capacitor based multiplier. Subsequently, the scaled maximum values are transferred (907) to a second switch-capacitor based adder, wherein the second switch-capacitor based adder and the switch-capacitor based multiplier perform the convolution operation. The ReLU operation is employed (908) on an output of the second switch-capacitor based adder, and an output of the hidden layer is transferred (909) to a decision layer. At the decision layer, the output is classified (910) based on one or more predefined threshold values, by an exponential and divisional implementation module."
Disclaimer: Curated by HT Syndication.
