Small Object Detection with Multiscale Features

Table 1

The process of model training.

Training process
Input: VGG_CNN_M_1024 and image
Output: detection model
Step 1 Initialize the ImageNet pre training model VGG_CNN_M_1024 and train the RPN network.
(1) Initialize network parameters using pre training model parameters
(2) Initialization of caffe
(3) Prepare for roidb and imdb
(4) Set output path to save the caffe module of intermediate generated.
(5) Training RPN and save the weight of the network
Step 2 Using the trained RPN network in step 1, we generate the ROIs information and the probability distribution of the foreground objects in the proposal regions.
Step 3 First training Fast RCNN network
(1) The proposal regions got from step 2 are sent to the ROIs
(2) The probability distribution of foreground objects is sent to the network as the weight of the objects in the proposal regions
(3) By comparing the size of Caffe blob, we get the weight of objects outside the proposal regions
(4) The loss-cls and loss-box loss functions are calculated, classify and locate objects, obtain the detection models.
Step 4 Replace the detection model obtained in step 3 with the ImageNet network model in step 1, repeat steps 1 to 3, and the final model is the training model.