Keywords

1 Introduction

Object detection is the process of finding instances of semantic objects such as persons, animals and vehicle in images or videos. One way to address the object detection problem is to use image segmentation methods. Currently, many applications require an automatic segmentation step in very different fields, with specificities related to the processed images [1]. Image segmentation is an essential process for many applications, such as object recognition, target tracking, content-based image retrieval [2]. Image segmentation methods are mainly grouped into three categories: Threshold-based techniques, Edge-based techniques and Region-based techniques [3].

More recently, hybrid models, which incorporate both edge and region information in their segmentation algorithms, have been introduced [4]. Image segmentation can be classified into supervised and unsupervised segmentation methods [5]. In this work, we are interested in unsupervised methods (automated segmentation methods).

The unsupervised segmentation algorithms divides the image into homogeneous regions (a set of neighboring pixels) but without considering the significance of these regions. While human segmentation tends to divide the image into objects that have a meaning (called semantic segmentation). For this reason, researchers try to bring automatic segmentation as close as possible to human segmentation.

In this paper, we propose a fully automated approach to region segmentation in images, that combines region growing and edge preservation methods. We also give a brief description of our implementation and evaluation results obtained using the Berkeley Segmentation DatasetFootnote 1 (BSD).

2 Related Works

Object detection can be treated in two phases. The first phase uses the region proposals method to segment an image into candidate objects. The second phase classifies every proposal region into different classes of objects using trained classifiers such as SVM [6] or Neural Network [7]. In object detection, some of the efficient techniques exploit sliding window and boosting [8]. The main drawback of the sliding window is that the number of proposals can have a complexity of \(\text {O}(10^{6})\) for a 640\(\,\times \,\)480 image [9], which increases the computational time and cost. To solve these problems, the authors of [9] proposed to use the superpixel labeling method to effectively detect objects. Superpixel represents a set of pixels which have similar colors and spatial relationships. Unfortunately, the use of superpixels also suffers from the problem of choosing the number of clusters K (maximal number of objects present in an image). A small value will generate a sub-segmentation and a great value will generate an over-segmentation.

There are several image segmentation techniques that were proposed in the literature. They are classified in three categories: Threshold Technique Segmentation, Region Based Segmentation and Edge Based Segmentation. Each one of these approaches has advantages and disadvantages depending on the application domains [10, 11]. Recently, the authors of [1] proposed an interested work based on region merging strategies that start by an over-segmentation process. Unfortunately, the author does not quote the execution time of the segmentation process.

We present in this work, an unsupervised image segmentation method for object detection. The proposed approach uses the region growing method combined with edge detection and some filters like bilateral filter, which serves to smooth images.

3 Proposed Method

In this work, we propose a new hybrid segmentation method that combines two techniques: edge based region growing and region merging method for object detection. For this reason, we use our algorithm which is based on region growing method (pixel aggregation) reinforced using canny edge to preserve the boundaries of objects. The generic algorithm is composed of three steps. we present the steps of our method Fig. 1 which are detailed in the following subsections.

Fig. 1.
figure 1

Architecture of proposed method

Full size image

3.1 Image Pre-processing

  • Denoise image using median filter,

  • Compute canny edge for original image I,

  • Smooth original image I (using bilateral filter),

  • Convert smoothed image to LAB color space,

  • Normalize color values between 0 and 1.

3.2 Region Growing Segmentation

  • Select initial seed point located in position (1,1) from image I,

  • Define a similarity measure S(p,q) based on two criteria:

    • Euclidean distance between the LAB color of pixels p and q,

    • Verify non-edge pixels for pixel p and q.

  • Evaluate the neighbors of seed point \((\text {seed}_{p})\) as above:

    • If the neighboring pixels of seed point satisfies the defined criteria, they will be grown (add pixel q to \(\text {seed}_{p}\)). The 3 neighbors of pixel q are added into the neighbors list of p according to the moving direction between pixels p and q,

    • Else consider pixel q as a new seed point \((\text {seed}_{q})\).

  • Repeat the process until all pixels in the image are treated.

We use 8-connected neighborhood to grow the neighboring pixels to the current seed. The formula of Euclidian distance used is:

$$\begin{aligned} D_{Lab}\left( x,y\right) =\sqrt{{(L_x-L_y)}^2+{(a_x-a_y)}^2+{(b_x-b_y)}^2} \end{aligned}$$
(1)

We grow each pixel that has distance value \(\text {D}_{Lab}\) less than or equal to the best threshold; I choose value 0.10 (that is perceptually acceptable). To examine the neighboring pixels, we enumerate each neighboring pixels of a given pixel p in a clockwise direction {0,1,..,7} which corresponds to the orientation degree of pixels. This technique improves the processing time and allows a fast scan of the whole image.

3.3 Region Merging

For an initial-segmented image:

  • Define a criterion for merging two adjacent regions,

  • Compute features for each region and construct a region adjacency matrix,

  • Iteratively merge all adjacent regions satisfying the merging criterion,

  • Repeat process until all regions are merged.

The merging criterions that we propose are the following:

  • Regions are allowed to merge if they have small color differences (using previous defined Euclidian distance and mean colors of each region),

  • Region that have small size will be merged with the biggest adjacent region,

  • All adjacent regions with Lambda value less than a given threshold will be merged.

We use Full Lambda Schedule method introduced by Robinson [12], described as follow:

$$\begin{aligned} L\left( v_i,v_j\right) \mathrm {=}\frac{\frac{|{\mathrm {O}}_{\mathrm {i}}|. |{{\mathrm {O}}_{\mathrm {j}}}|}{|{\mathrm {O}}_{\mathrm {i}}|+ |{{\mathrm {O}}_{\mathrm {j}}}|}{.||u_i-u_j||}^2}{l(v_i,v_j)} \end{aligned}$$
(2)

where:

figure a

4 Experimental Evaluation

In order to make an objective comparison between different segmentation methods, we use some evaluation criteria wich have already been defined in literature. Briefly stated, there are two main approaches [13]: supervised evaluation criteria and unsupervised evaluation criteria. Supervised evaluation use a ground truth, whereas unsupervised evaluation enable the quantification quality of a segmentation result without any prior knowledge [14]. The evaluation of a segmentation result makes sense at a given level of precision.

In our work, we choose to use a supervised evaluation. Our evaluation is based on the Berkeley Segmentation Dataset and for each of these algorithms, we examine two metrics:

Table 1. Quantitative comparison of different algorithms
Full size table
Table 2. Best results of segmentation process
Full size table
  1. 1.

    Probabilistic Rand Index: measures the probability that the pair of samples have consistent labels in the two segmentations. The range of PRI is between [0,1], with larger value indicating greater similarity between two segmentations [15];

  2. 2.

    Variation of Information: measures how much we can know of one segmentation given another segmentation [15]. The range of VoI is between \([0,\infty {}]\), smaller value indicating better results.

Our experiments were performed on a machine with an intel core i7-2630QM. We compare our method with other segmentation algorithms, including: Efficient Graph-Based Image Segmentation [16], Texture and Boundary Encoding-based Segmentation [17], Weighted Modularity Segmentation [18], Mean-Shift [19], Marker Controlled Watershed [20], Multiscale Normalized Cut [21]. The segmentation results are presented in Table 1.

Our method for unsupervised image segmentation gives acceptable results especially we have used only color feature for similarity measure. We observe that the average time execution is less than 25 s. As a result, we can say that our method gives a high quality segmentation result with minimum time consumption. Table 2 shows some examples:

5 Conclusions

In this paper, we have proposed a hybrid method for unsupervised image segmentation. We have taken advantage of the fast edge based region growing algorithm and region merging techniques to overcome the problems of over-segmentation and sub-segmentation encountered in the super-pixel and sliding window for object detection.

Our algorithms have been tested on the publicly available Berkeley Segmentation Dataset as well as on the Semantic Segmentation Dataset. Then, they have been compared with other popular algorithms. Experimental results have demonstrated that our algorithms have produced acceptable results. However, the obtained regions do not exactly match the shapes of the objects contained in the images. Other treatment would be required to make the correspondence between regions and objects. Thereby, we can confirm the complexity of the segmentation domain and then we will explore other approaches to improve the object detection process by using machine learning in our future works.