HSPOG: An Optimized Target Recognition Method Based on Histogram of Spatial Pyramid Oriented Gradients

HSPOG is an improvement of HOG. For the same target in two images with different sizes, the two target image areas would be divided into $2^n$ cells, and the relative characteristics between corresponding cells of the two areas would remain unchanged even if the cells have different sizes. The goal of HSPOG is to extract stable target features from different image targets and remove the unique image target size and ratio requirements of HOG. HOG always scales an image target into a fixed size of $64\text{pixel}\times 128\text{pixel}$ and defines the cell size of the target as $8\text{pixel}\times 8\text{pixel}$ . However, we think that the cell size should be dynamically determined by the cell number. So when calculating HSPOG features, we scale a target image dynamically without distorting (see Fig. 1 ). The scale ratio can be calculated by Eq. ( 1 ),

(1) $$\mathrm{ratio}=\mathrm{A}\mathrm{\_}\mathrm{scale}/\min (\mathrm{row},\mathrm{col})$$

where $\mathrm{A}\mathrm{\_}\mathrm{scale}$ is the adjustment scale.

10.26599/TST.2020.9010011.F001 Fig. 1Cropping or warping to fit a fixed size.

To get more target features of pyramid scales, the bigger the value of n in principle, the higher the accuracy of recognition. However, each increment of n by 1 will exponentially increase the computation amount, so we make $2⩽n⩽5$ empirically. There is a feature vector for each scale, adding all the features together can form a long descriptor, which includes coarse features on large spatial scales and fine features on small spatial scales. HSPOG is described in Algorithm 1. The length of the small edge of an input image after zooming may be 64, if we specify the cell number along the small edge to be 8, the parameter n is 3. If the large edge of an image cannot be divided exactly by $2^n$ , we must fill the image by zeros before it can be divided exactly by $2^n$ , as shown in Fig. 2 .

10.26599/TST.2020.9010011.F002 Fig. 2Zooming method used in HSPOG. If the large edge of an image can not be divided by 2 ${}^n$ , it would be filled with zeros along the large edge until it can be divided by 2 ${}^n$ exactly.

The flow chart of HSPOG is shown in Fig. 3 . HSPOG can avoid image distortion and information loss. The size of an image cell is calculated dynamically, while the corresponding of HOG is constant. Dynamic cell determination ensures the ratio of the cell proportions the same. HSPOG contains multiple spatial pyramid scales for feature calculation, which enables HSPOG to integrate coarse features at large spatial scales and fine features at small spatial scales while preserving information integrity.

10.26599/TST.2020.9010011.F003 Fig. 3Flow chart of HSPOG ( f is the feature vector of HSPOG).

The rapid detection and recognition of an image target is key to engineering applications. The article improves the algorithm of HSPOG to accelerate the calculation of features.

When calculating a feature map, HOG divides 180 degrees into 9 oriented blocks commonly, and the degrees between 180 and 360 can be considered as its negative situations. HSPOG not only divides 180 degrees into 9 blocks, but also divides 360 degrees into 18 blocks to obtain better oriented gradient features. In this process, trilinear interpolation^{[ 15 ]} is no longer used to calculate the degrees of its adjacent fans.

In HSPOG, we use bilinear interpolation instead of trilinear interpolation to calculate the weights of each pixel to neighboring cells, which is a fast calculating method^{[ 16 ]}. For example, the similarity between 85 and 80 degrees is 0.75, and the similarity between 100 and 85 degrees is 0.25. So, the best position of 85 degrees may be determined as the fifth block. We illustrate the processing in Fig. 4 . Feature map of each cell will be normalized four times because it may be contained in four blocks. When all pixel features are added to HSPOG together for an image target, we will discard the features of boundary pixels because they cannot be normalized four times. For an image target, assuming $n_{\mathrm{row}}$ is the number of cells in the row and $n_{\mathrm{col}}$ is the number of cells in the column, the length of the image target HSPOG is $(n_{\mathrm{row}}-2)\times (n_{\mathrm{col}}-2)\times 32$ . The 32 features in the previous equation contains 9 containers (see Fig. 4a ), 18 containers (see Fig. 4b ), 4 texture features, and 1 truncated feature, fast HSPOG is shown in Algorithm 2. The acceleration algorithm is an approximate simplification of the original algorithm, but the attenuation is negligible.

10.26599/TST.2020.9010011.F004 Fig. 4Two types of bins ( ${\mathit{Z}}_{\mathit{i}}$ is the i-th bin).

The spatial pyramid scales shown in Fig. 5 will be used when computing the HSPOG for an image. The scales play important roles in conventional methods, e.g., the Scale-Invariant Feature Transform (SIFT) vectors are also collected at multiple scales^{[ 7 , 17 ]}. So, we also compute the HSPOG at multiple scales, in the processing, scales include $4\times 4$ , $8\times 8$ , and $16\times 16$ . All of these pyramid scales make HSPOG more potent during the processing of object recognition missions.

10.26599/TST.2020.9010011.F005 Fig. 5Spatial pyramid scales and the feature map of an image.