Mastering Computer Vision with TensorFlow 2.x

GoogLeNet detection

The inception network (also known as GoogLeNet) improved upon the two-stage layer (region proposal based on color, texture, size, and shape, and then CNN for classification) proposal for recurrent convolutional neural networks (R-CNN).

First, it replaced AlexNet with an improved inception for CNN. Next, the region proposal step is improved by combining the selective search (in R-CNN) approach with multi-box predictions for higher object bounding box recall. The region proposal is reduced by about 60% (from 2,000 to 1,200) while increasing the coverage from 92% to 93%, resulting in a 1% improvement of the mean average precision for the single model case. Overall, the accuracy improved from 40% to 43.9%.