Heterogeneous convolutional neural network for multi-problem solving
A heterogeneous convolutional neural network (HCNN) system includes a visual reception system generating an input image. A feature extraction layer (FEL) portion of convolutional neural networks includes multiple convolution, pooling and activation layers stacked together. The FEL includes multiple stacked layers, a first set of layers learning to represent data in a simple form including horizontal and vertical lines and blobs of colors. Following layers capture more complex shapes such as circles, rectangles, and triangles. Subsequent layers pick up complex feature combinations to form a representation including wheels, faces and grids. The FEL portion outputs data to each of: a first sub-network which performs a first task of object detection, classification, and localization for classes of objects in the input image to create a detected object table; and a second sub-network which performs a second task of defining a pixel level segmentation to create a segmentation data set.
1. A heterogeneous convolutional neural network (HCNN) system, comprising:
a feature extraction layer (FEL) portion that receives an input image, the FEL portion conducting a learning operation to learn to represent a first stage of data of the input image and outputting the first stage of data to both of:
a first sub-network directly receiving the first stage of data from the FEL portion and performing a first task; and
a second sub-network directly receiving the first stage of data from the FEL portion and performing a second task,
wherein the first stage of data comprises a first feature map,
wherein the first sub-network includes a first convolution and pooling layer (CPL) portion receiving the first stage of data and capturing a second stage of data, and
wherein the second stage of data includes a second feature map different than the first feature map.
2. The HCNN system of claim 1 , wherein the second stage of data includes a second feature map different than the first feature map.
3. The HCNN system of claim 2 , wherein the first sub-network includes a second convolution and pooling layer (CPL) portion receiving the second stage of data and capturing a third stage of data.
4. The HCNN system of claim 3 , wherein the third stage of data includes a third feature map different than the first feature map and the second feature map.
5. The HCNN system of claim 4 , wherein the second stage of data and the third stage of data are both input into the second sub-network.
6. The HCNN system of claim 4 , wherein the second sub-network includes a third convolution layer and pooling payer (CPL) portion, a fourth convolution layer and pooling payer (CPL) portion, and a fifth convolution layer and pooling payer (CPL) portion.
7. The HCNN system of claim 6 , wherein the third CPL portion receives the first stage of data, the fourth CPL portion receives the second stage of data, and the fifth CPL portion receives the third stage of data.
8. The HCNN system of claim 7 , wherein outputs from the third CPL portion, the fourth CPL portion, and the fifth CPL portion include pixel level segmentation and are combined into a segmentation data set.
9. The HCNN system of claim 4 , wherein the first stage of data, the second stage of data and the third stage of data are input into a fully connected layer that determines multiple confidence levels for an object identified by the first stage of data, the second stage of data, and the third stage of data.
10. The HCNN system of claim 9 , wherein the fully connected layer communicates the multiple confidence levels to a non-maximum suppression module that reduces the multiple confidence levels to a single confidence level for the object.
11. The HCNN system of claim 1 , wherein the first task includes object detection, classification, confidence, and localization for objects in the input image to create a detected object table.
12. The HCNN system of claim 1 , wherein the second task includes defining a pixel level segmentation to create a segmentation data set.
13. A system for a vehicle, the system comprising:
a visual reception system generating an input image;
a heterogeneous convolutional neural network (HCNN) system, comprising:
a feature extraction layer (FEL) portion that receives the input image, the FEL portion generating a first stage of data of the input image and outputting the first stage of data to both of:
a first sub-network directly receiving the first stage of data from the FEL portion and performing object detection, classification, confidence, and localization for objects in the input image to create a detected object table; and
a second sub-network directly receiving the first stage of data from the FEL portion and defining a pixel level segmentation to create a segmentation data set, wherein multiple stages of data generated in the first subnetwork is input into the second sub-network,
wherein the first stage of data comprises a first feature map,
wherein the first sub-network includes a first convolution and pooling layer (CPL) portion receiving the first stage of data and capturing a second stage of data, and
wherein the second stage of data includes a second feature map different than the first feature map.
14. The system of claim 13 , wherein the multiple stages of data generated in the first sub-network include multiple feature maps.
15. The system of claim 13 , wherein a second convolution and pooling layer (CPL) portion receiving the second stage of data and capturing a third stage of data, and the second stage of data and the third stage of data are both input into the second sub-network, wherein the second stage of data includes a second feature map different than the first feature map and the third stage of data includes a third feature map different than the first feature map and the second feature map.
16. The system of claim 15 , wherein the second sub-network includes a third convolution layer and pooling payer (CPL) portion, a fourth convolution layer and pooling payer (CPL) portion, and a fifth convolution layer and pooling payer (CPL) portion, wherein the third CPL portion receives the first stage of data, the fourth CPL portion receives the second stage of data, and the fifth CPL portion receives the third stage of data.
17. The system of claim 16 , wherein the first stage of data, the second stage of data and the third stage of data are input into a fully connected layer that determines multiple confidence levels for an object identified by the first stage of data, the second stage of data, and the third stage of data, and the fully connected layer communicates the multiple confidence levels to a nonmaximum suppression module that reduces the multiple confidence levels to a single confidence level for the object.
18. The system of claim 16 , wherein outputs from the third CPL portion, the fourth CPL portion, and the fifth CPL portion include pixel level segmentation and are combined into the segmentation data set.