Page 88 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)
P. 88

72


                     2.3   Measurements

                     In  this  paper,  after  experimentally  training  the  model,  we  used measurements  to
                     compare, evaluate, specifically:


                     MAP (Mean Average Precision): This is a measure of the aggregate results of many
                     queries  applied  to  the  search  system.  To  calculate,  we  must  have  AP (Average
                     Precision) as the average of the precisions at the threshold points returned by each
                     correct result, written with the following formula:






                     with recalls(n) = Rs(n) = 0, precisions(n) = Ps(n) = 1, n = threshold coefficient.
                       Once AP is available, the formula for mAP is written as follows:




                                           AP k = AP value of class k, n = number of layers

                     F1 curve: is a tool commonly used in classification tasks to evaluate the performance
                     of a model with the formula:
                                                                                                     (3)



                                                    Table 3. Measurements
                         Measurability                  Meaning                        Formula
                      mAP                 Average accuracy measurement
                      Precision           Prediction accuracy                        TP/ (TP+FP)
                      Recall              The goodness of correct predictability     TP/(TP+FN)

                      Training time       Training period
                      Memory              Memory used


                     Where, TP is the rate of true positives, TN is the rate of true negatives, FP is the rate of
                     false positives, and FN is the rate of false negatives.

                     2.4   Results

                     Object Detection has color image input and output is object and position of objects in
                     the  image.  Real-time  object  detection  with  high  accuracy  is  a  requirement  for  the
                     problem  of  identifying  hot  weapons,  specifically  in  this  article  is  Pistol  (pistol)
                     identification.  During training, if the error level (Loss) of the current loop and the
                     average error of the model (Model Avg Loss) change little, stop the training process.
                     Testing result was shown on some images containing pistol (pistols) and showed that
                     the objects detected in the images were rectangles with red borders, the results of the




                     CITA 2023                                                   ISBN: 978-604-80-8083-9
   83   84   85   86   87   88   89   90   91   92   93