Hello,

today I had trained CNN (Convolutional Neural Network) on train dataset containing two thousands images (1000 images with people on them and 1000 negative examples). I used images from "COCO dataset" - see link:

https://cocodataset.org/#home

I initialy used program written by me with "OpenCV" library to split examples into two categories "Men on photo" and "No men on photo" and then corrected mistakes manually. Then I used "Google Colboratory" account to define CNN network with Keras framework and train it on this dataset prepared in advance. The prediction results after training this CNN are very encouraging and much better than using "HOG descriptors and SVM" algorithm with OpenCV. Here is sample prediction on ten images which CNN has never seen before:


[0. 1.] label means "no men" and [1. 0.} means "is men on image". Tomorrow I will examine this results more accurately.

Best regards

 

Hello,

today I achieved level of correct prediction of CNN network equal 99.55 percent on training data and 94.14 percent for data that CNN had seen for first time. It is not too easy to achive such level of right predictions. Below screen-shot from "Jupyter notebook" (from my Google colaboratory account with Keras framework) with results for last phase of training this neural network:


I will write more about it later in separate post.

Best Regards